They slashed pricing, it's now nearly half vs the preview (Input: $0.7/M Tokens, Output: $2.1/M Tokens vs Input: $1.36/M Tokens, Output: $4.18/M Tokens).
At the original pricing in my evals, Mistral Large 4 Preview was not competitive. With this price point and some improvements on top, this full release warrants a major reassessment.
They didn't. Original price was maybe higher but they slashed it during preview and in fact it was lower by a bit ($0.68/M tokens) and $2.09 for output.
I've been using it for my personal project and it's still too expensive and very slow when thinking. I think we need one more generation from mistral and with maybe 50% of this already "slashed" price for this to be competitive (I use Opus 5.5 at work and it's still much better than this).
I just checked, seems you are right. They had roughly a week with the lower pricing, then increased it, now decrease it again. Very odd. My testing was with the higher pricing and looking at AA, theirs seems to have been too.
Token providers in aggregate work a bit like airlines so the end pricing should also work a bit like airlines; expect budget tokens, fast boarding tokens, window seat tokens, last minute tokens, miles for tokens, … and naturally prices fluctuating by the hour.
Not sure it’s particularly odd. Several models of late have launched with an introductory promotional pricing, only to later lock in the promotional pricing.
Too many models now to remember specifics - I believe deepseek v4 did this and one of the gpt models (5? 5.1? 6?) did this if I recall correctly. I’m also fairly confident Google has had a few promotional periods on launch of the model, I seem to remember them starting the trend.
Yeah I tried mistral 4 when it came out and it honestly felt like going back in time 6 months, suddenly my agents were making all kinds of mistakes that I had kind of forgotten agents regularly used to make
> They slashed pricing, it's now nearly half vs the preview (Input: $0.7/M Tokens, Output: $2.1/M Tokens vs Input: $1.36/M Tokens, Output: $4.18/M Tokens).
Those in the EU wanting EU-only inference should be aware Mistral put a 10% tax ("upcharge") on Regional Inference[1].
Given the only practical alternative to a jurisdiction with strong data protection laws is a home rig, as the saying goes, still cheap at twice the price.
Given the size of the model, that's very aggressive pricing.
I wonder if it is serving at a loss to get training feedback (maybe using the margins they get from GLM 5.3 serving which is a slightly smaller model but double the price).
based on Anthropic's S1 no one is serving models at a loss, but cutting your price to keep usage high makes a lot of sense when you have your inference booked for long periods.
> this full release warrants a major reassessment.
I really don't think it does, the benchmark numbers are dreadful and the cost is higher than Sol. If you have to use an EU model knock yourself out but there are a ton of models that are a lot smarter for a lot less money.
So if scores are good, they must be benchmaxxed, and if they are bad, they must not be benchmaxxed. So basically, the narrative is whatever you want it, no matter what the data says.
No, if a model scores good on benchmarks it could be that they have benchmaxxed, if a modern model scores bad on benchmarks it could be a really bad model but it almost certainly isn't benchmaxxed. It's relatively easy to make a model do well on particular benchmarks, companies that don't care about that stand a better chance at making a model that does well in general because tweaking a model for benchmarks usually is at the expense of something else.
I run my own benchmarks (this model is - unfortunately too large so that won't be happening here) on data that no outsider has ever seen and this gives sometimes results that are not in line with the order that you would expect if you took the various benchmark sites into account. It is of course entirely possible that our benchmark is not representative but it seems to work quite well as a predictor for which models have value for us and which do not. Recently we looked a model that has been pushed with a significant amount of hype and for the life of me we could not get it to perform even close to the published benchmarks, whereas the models we used for controls were definitely within the margin of error. It's not always equally cut-and-dry.
I don't quite believe the benchmarks, especially supplied by the company.
Qwen3.8-Flash-Next consistently benchmarks better than both GLM5.3-Flash and the previous DeepSeek v4.1 Flash 0731. The DS 0731 was so atrocious it was getting beaten by small Nemotron in my coding tests (subset of tasks from open source coding benchmarks).
Now, new DeepSeek V4.1 Flash 1017 really gets the numbers they promised for the first one, but it's too big for me to run the 4 bit quant locally at interactive speeds.
Irregardless of wether or not Mistral Large 4 is competitive or not, I do not trust the AA numbers anymore, for various reasons. Them changing their numbers around so much recently brings the entire validity of this in question, not to mention if such a combined number is even important at all. The combined "intelligence" score of AA includes areas I do not care about or need at all.
Agree. They had just fiddled with the formula a few weeks prior to this, too. There seems to be an effort to get a certain result from the model rankings.
Exactly, these general scores are pointless. Let's say the model is amazing on A but crap on B and C. It gets low score, but you still would benefit greatly if you need it for A.
"Cyber benchmarks, while achieving an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models."
I feel like stating the obvious here and hence like I'm missing something. Where is the statistic about not refusing the non malicious requests? I can achieve the same performance with refusing when random(1000) < 967.
These statistics and safeguards are pretty much meaningless, because it's completely trivial to get around them. You can prefill and edit the model's responses/thinking and force it to agree to virtually anything even with unmodified weights, and if you don't want to bother with jailbreaking and can host it yourself it's essentially a one-man weekend project to uncensor the model and have it agree with anything you want by default.
Guardrails on open-weight models is security theater purely to shield whoever released them from a potential PR nightmare.
It's unfortunate, because the security theater performed here detracts from the model's overall cybersecurity performance. And uncensoring has another small hit to performance.
Much like having a hacking tool doesn't mean you're doing anything illegal, attacking a machine doesn't necessarily mean you're committing a crime. I can hack the living shit out of my local network if I want to.
"The model failed to refuse a cyber request when it should have" != "the agent hacked someone". For example, refusals also apply to just asking the model how to do something, without extending it any agency.
They were given day zero access to release their quantizations[0]. The lowest quant will still require 256GB VRAM to run inferences, but as always I'm amazed by the Unsloth team by how much they can reduce the requirements for these powerful LLM's, and the easy-to-follow guides to get started with using them.
Oh, that's a tough one. Do you have a fast drive? If you do you might try a 3 to 5B active MOE model, use the rest as cache to stream weights from NVMe. That would be your best bet. You'll be in the 'optimization league' ;)
There are two big boards on that page, there is a four PCIe slot board and a five PCIe slot board. The spacing on the '4' is a lot wider than on the '5' (it is also non-standard so a normal enclosure will not work).
That price was valid for multiple parts of it as late as a week or two ago (I am not checking these prices daily).
The RTX PRO 4000's can still be had at the same price I paid for them, ditto the switch backplane, the enclosure, the PSU and the 2nd hand 3090's as well as the 4090's. What has gone up (considerably) in price is the regular RAM.
It's interesting how all the recent SOTA models are in the 1T-3T range. This even has a large number of active parameters (A52B). It is also likely that OpenAI, Anthropic, et. al. have similar sized models.
This has two effects: 1) it prevents effective use of these models on consumer-grade hardware; and 2) it pushes the requirements to train and run these models on existing and future data centres, making them more expensive to run.
It would be nice to have more slimmer models and more active research that isn't just "let's throw more compute at the problem". Maybe they'll train smaller models from the large model like Qwen have done in the past.
I'm not sure it really matters. I think a quantized version of a large model ends up being only slightly worse than a model trained at a smaller size to begin with. And like you point out, large models can also be used to distill smaller models, so it's good to have them out there.
The hardware is certainly not cheap, but you can put 512gb of vram on your desk if you are willing to throw $30k at Apple, which is a manageable expense for a lot of small businesses. Bridging a few devices is one way to run large models, r ablation and quantization can do a lot too. You definitely need a data center if you aim to make the model directly accessible to the public, with whatever fine tuning or secret sauce you add. If you want to run it in-house as a coding engine or research tool you can likely reduce its footprint substantially.
I think a small business, even a medium sized one, is better served by either renting proper hardware in a cloud or finding an inference provider to sign a contract with.
I bet you could get the same performance with smaller models, but as english-only, by removing training data in non-english (keep code, translate in-code strings mechanically prior)
I doubt multi-lungualism is taking up that much space in the weights. Languages are systematic, and most of them share a lot of features and can be compressed into the same representation together.
The real problem is all the random facts about the world that are not easily compressible, and you want the LLM to know a lot of random facts.
AFAICT there are several things that affect the size of the model:
1. the number of transformer layers -- e.g. each layer adds the ability to provide more information to the residual stream [1], [2]; each layer is `tokens * features * active context`
2. the number of tokens -- the more tokens there are the more word fragments and Unicode characters you can represent
3. the number of features assigned for a token -- these are used to associate trained semantic information for the token; with the active context this can encode contextual information for that token (the idea being that the model can differentiate senses of a word like "bank" depending on its context, so "bank (river)" will have different weights than "bank (building for money)")
4. the active context -- this is the number of `token * features` to include when predicting the next token
One of the key features of LLMs and transformer architectures is that training across different languages helps the model learn the underlying concepts and associations, so training on English only would lose that.
A better approach may be something like a byte architecture or a model where the tokenization is trained with the model (similar to vision and audio) instead of as a separate step. Or maybe training smaller models longer to take advantage of double descent [3], [4].
This is strategic. By making it so consumer grade hardware can't serve it, you have to pay them for access to it, and that funds the server hardware market, which keeps the AI market afloat. If they made smaller models, we would all be running local AI in a few years, nobody would pay them, and they would go bankrupt (and a worldwide recession would begin). It's going to be up to us (FOSS community) to train and publish our own small models. (Also because the US and China are going to make releasing open weights illegal or highly regulated soon)
I don't think they're purposefully bloated, any more than literally all software today is. It all is bloated, but I don't think anyone set out for that. Despite that, the bloat is now the business case. If one of these AI CEOs announced "our main goal is to make tiny models for consumer devices", their stock would crash.
Mistral has been great for us as a provider for our EU-based company. We use Mistral Large 4, GLM 5.3 Flash and DeepSeek v4.1 Flash 1017 through their API for our B2B2C app for complex work around building blocks for documents and custom tools. A few months ago we weren't sure which EU provider to use, but since they introduced 3rd party models, added more features (now Responses API etc.) and improved their own model, we are quite happy-
> but since they introduced 3rd party models, added more features (now Responses API etc.) and improved their own model, we are quite happy-
Yeah, serving GLM is a pretty good play on their side. I know they have been criticized for it ("so, you're just reselling GLM...") but it seems practical as in it helps EU companies use it.
I have absolutely no problem with them hosting and selling GLM as an offering, it just makes sense for them to do this as an option. As a long time subscriber however, I don't like at all that they have started using GLM as the default model in the Vibe app (formerly Le Chat), with seemingly no way to switch to Mistral's own model. I'm paying a monthly subscription to Mistral explicitly to use a European tool in my daily non technical LLM needs. This feels off and I'm reconsidering my subscription. If all they are giving me is a hosted Chinese model, I can probably get that cheaper elsewhere.
So they can use GLM from a trusted provider, that is highly committed to respect local laws and regulations.
Just think about a defense contractor. Choosing trustworthy vendors is crucial to them. They can't just use a random provider from Open Router. They need to vet and certify their vendors. If they already have Mistral on the list of approved vendors, they can now just switch to the most recent open weight LLM if Mistral starts serving it.
They can use the model and the data and requests all stay in the EU region. As opposed to sending to Anthropic, Microsoft or Google. Granted, those companies can assert that data will stay in the EU region too, but they'd have to be believed, which, some may find tricky lately.
In this case it's a EU company serving either an EU built model or a Chinese model. For the Chinese models they are not forwarding the requests to China, the model is still running in EU.
GCBD operates the iCloud China service with Apple, and I imagine Apple, Inc and its associated LEA portal/systems does not have access (otherwise there's no chance China would let iCloud China be a thing).
This is also where the words of law and pragmatic reality intercede.
> ... As opposed to sending to Anthropic, Microsoft or Google. Granted, those companies can assert that data will stay in the EU region too, but they'd have to be believed, which, some may find tricky lately
Sovereignty does not always require self-hosting; sometimes it just requires deployments that meet your control requirements. For organizations that prefer a managed service, Mistral Large 4 is available through Microsoft Foundry, with private deployment options for Azure developers.
There's many EU companies that want to have a contractual relationship with only another EU company and be absolutely assured that the models, the contents of the KV cache, any logging, all access control/credential systems, billing platform, etc are all hosted on EU-company owned hardware located in an EU jurisdiction.
There is a similar thing happening now on a smaller scale for Canadian companies that want to migrate off USA-based platforms for all sorts of SaaS and reasons.
I didn’t know mistral started serving deepseek-flash 4.1! This is awesome. A little more expensive than most providers though. $0.03/$0.007/$1.20 on mistral vs deepseeks own peak pricing of $0.03/$0.006/$1.20 (vs deepseeks off peak pricing of $0.015/$0.003/$0.60).
I prefer using a European provider though so this is actually cheaper than I can find elsewhere.
Another vote here - moving a lot of our stuff over to Mistral. Very much happier keeping data in Europe rather than leeching it all over the US. Plus, the orange lunatic.
> 2. If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose. The Licensee may contact Mistral, using this form: https://mistral.ai/contact/, to request a commercial license. [0]
Calling this "open weight" is misleading. That's why we had huge backlash when Redis, Elasticsearch, and Terraform went source-available, right?
Weight-available is a better adjective for this kind of licensing. Same for Qwen, GLM, Kimi, MiniMax, etc. If they really want to be open source, they should pick an OSI-approved license.
Training a frontier model is prohibitively expensive, so it’s understandable that open weights aren’t truly open source. In fact, I worry whether open weights will even survive[1] once the current AI cycle cools down. An even bigger issue is that as long as model development remains this capital intensive, a genuinely open-source community around frontier AI is virtually impossible. That will have massive implications for our careers. The more capital concentrates at the top, the less leverage we’ll have in the job market.
[1] Arguably hardware companies like Nvidia have incentives to subsidize open models to drive GPU demand. That said, I'm not sure how sustainable and sufficient that incentive can be.
No, they're still capable of calling it something different, instead of implying open source licenses that they aren't actually using. That's a valid complaint that's kind of unrelated to any of this other stuff.
Whether it's feasible to release everything under open terms is actually a different issue from whether they're being truthful in how they're advertising it.
> Arguably hardware companies like Nvidia have incentives to subsidize open models to drive GPU demand
Not speaking on behalf of my employer and I have no insight into this strategy at all, but I feel like the sale of consumer GPUs is so small by comparison to data centers (more than 10x from what I can tell online) that more users of open weights wouldn't likely make much difference in the earnings report.
There’s no reason to believe that companies running open weights models would use consumer hardware. I can speak on behalf of my employer and say that we’d vastly prefer to have enterprise grade servers running inference on premises.
Not consumer demand. I was assuming that it's Nvidia's interest that intelligence is not only in a couple of frontier lab's hands, or they can choke the demand to suit their needs. If any company can access a frontier model without going through the frontier labs, the ecosystem will be larger and healthier for everyone, including Nvidia.
Open weights models in the data center are more important than on consumer devices.
In the future, even consumers will launch weights and fine tunes they own in the cloud.
People don't own the fiber to their home, so why should they expect to run the GPUs? If they want smart coding and video models, they'll need H200s and B200s, and they'll need to rent them.
counterpoint: I have smart coding at this very second, and I am running it on 128GB of ddr4 with a 16GB gpu. (Qwen 3.8 Flash Next).
I expect to continue to have this, and more, as smaller models become even more capable. I also expect hardware prices to resume going down once the supply crunch ends.
I see no need to stop owning things or change my expectations around ownership, nor will I accept something like that.
> Arguably hardware companies like Nvidia have incentives to subsidize open models to drive GPU demand.
This does not mean that they would necessarily subsidise consumer demand. They may be very happy to sell palette-loads of GPUs to datacenter and large corporations.
Someone’s surely doing it atm we just don’t know yet. And then we scan all books but the stuff stays local. Perhaps possible, perhaps not. But I don’t see why everyone could not run given the epoch on his content only, reporting back findings (loss).
Well as long as we are all aware we’re hurdling towards techno communism. Better not fuck it up and end up ruled by fascists. Not looking great on that front so far.
> Either way, Mistral is a failing company with little relevance in the AI space and that's not going to save them.
Already ? I was wondering when some US based company will buy them including all that "soverignity" prompts :)
Local llms are what we need and to get that we need memory so we need computer builders from outside of US. Why ? Because when something external appears US companies also allows US citizens to have that too at lower price :) Not before external forces force them to be suddenly good hearted to their own peoples...
Can we also have coarser MoE's ? If I want to code I do not need llm be trained on past and present capitol names and chemistry and everything unrelevant to the task ! Watch Matrix again ! We need somewhat "inteligent" llms and knowledge can be uploaded just like Neo learned kung-fu :) - prompts and light tr^h^hweek of learning and some exams, we need kind of university for llms :) So llms finally fit into users computers... And when some grandmas or snowflakes start to whine tell them to learn - llms usage can be a skill finally ;)
Edit:
Wait, can we have something like interchengable weights so knowledge can be just downloaded to model folder ? Common weights format ? Reproducible builds ?
Their existence is the only thing keeping costs down across the board. If they had gotten open-weight model importing banned in the US a year ago, I'm sure they'd have 100% increased token cost relative to the advertised 100%+ intelligence improvements.
Open weight models are kind of weird because they're usually just a bunch of weights and maybe a chat template, not even test vectors or anything. Quite odd, and probably one of the reasons why vllm, sglang, the llamas etc. so often produce buggy output. It's actually quite strange how little attention is given to verifying these setups work.
Why cant the community piece together the open-weights using open community-based like, implementation or whatever magic sauce integrates those weights
It's not understandable for them to lie or mislead. Calling something open weights and not being remotely open in the commonly understood sense is misleading and not ok.
I don't think "you can do whatever you want other than just turn around and sell a chat wrapper for this" makes it immediately not open weights. Could it be more open? Sure. That doesn't mean it's misleading or a lie and I would strong disagree that it's "not being remotely open."
Open source doesn't mean "anyone can use the source code commercially". Open weights similarly doesn't mean the weights can be used for anything. It just means the weights are open for anyone to download and inspect.
> Open source doesn't mean "anyone can use the source code commercially".
Yes it does. It might have requirements that you release the code for any changes you make to it or link with it. But no OSI (or FSF) approved license prohibits you from using it commercially.
I'm a little confused by your distinction between vendor and lab. AI2 (Allen Institute for AI) is a western lab and they did release a fully open source model, with open weights, data, and training code (Olmo3).
We are now at the mercy of labs to provide frontier models as open-weight. Now that even some Chinese labs are releasing their models as weight-available, what stops them from prohibiting certain internal uses of new models? What about public chatbots?
Open models we have now will become obsolete within a few years. Imagine a world where our last open models were from the GPT-2 era.
I believe we need truly sovereign models, not in the Western sense, but from a community perspective. We need a model line that stays open-source across generations.
The problem is the massive amount of money, data, and talent to produce anything remotely competitive. Ideas have been floated about distributed training, but training requires insanely fast interconnect between nodes making it impossible with the current methodologies we have. Now if you mean funded at a national level, that seems more doable so long as the training runs are centralized in big DCs.
Great idea, let’s counter the power of a handful of large private organisations by putting that power in the hands of a single large government. What could possibly go wrong?
There are lots of governments around the world - some big, some small. Lots of discussions around digital sovereignty are framed in this manner - a model for Brazil, a model for Japan, etc. There certainly could be coalitions of governments or governments could strike out on their own. Whatever floats your boat.
There's always going to be some nerds out there that want to make and release a model as open, and share their knowledge along the way. I'm not too concerned.
The fear you're sharing has been shared since day 1 of open-weight models and it has yet to come to fruition.
Further, regardless of what any of these labs license their model as, the end user who has downloaded the weights can still use them for virtually any purpose (synthetic data, classification, etc) as much as they want and there is no way to reasonably enforce a license on those outputs.
So even if you can't run a service using this model's weights (without a license), the weights still do provide utility for training new models with any other license.
*Edit, responding to the replies: Constraints give rise to ingenuity. Who knows how we'll be training models in the future. Hopefully in a distributed fashion, maybe with some kind of consensus-driven training dataset selection. Crowd funding, etc. There's too many people that want this, and the knowledge on how to do it is out there (many times by now)
I think this is a quite optimistic view... releasing a SOTA model is a collosal capital investment, and unlike free software, it's not iterative - you can't make a small model and improve it one step at a time, you need to risk the whole load of cash, expertise and GPU time up front.
It's only happening now because labs have bucket tons of VC cash to burn and need a big public flex to raise more cash, or because they're subsidised by the Chinese government. I'm not massively confident either of those will stay true forever.
I think the problem is up to date knowledge isn't really the difficult or useful bit. We're sort of moving to a world where the real value in training comes from tool calling and performance over long context - your models shouldn't really be relying on their internal knowledge, but instead calling various tools. But the performance gap between now and six months for that is pretty huge.
Unless "some nerds" means "billionaires" - they may want to release new models as open source, but they will not have enough money to train a worthwhile model.
A completely different direction is to remove the fair use allowance for training on copyrighted information (unless you e.g. open source your weights).
We already had that world, the axis of evil stopped releasing open source models years ago. It’s only because of the Chinese that we kept getting them.
Of course they can. They're expensive to host, and if they're legally scrubbed off of places like huggingface then they are for all practical purposes 'closed' as in unavailable to people who don't already use them and if they're illegal or outpaced by Sota the ones who have copies can't even use them.
Thousands or more people already have copies and the license explicitly is irrevocable and allows redistribution. So if they take it off HF and HF is ok with that, people will just upload to anywhere else.
"outpaced by SOTA" doesn't matter at all, the models we have are plenty good enough for most real tasks.
If it is capital intensive and we require the work be done in the public benefit, the only options are international consortia or nationalization. Non-profits will be guided by the interests of the rich and are unaccountable to the public. Worker cooperatives are unlikely to be well capitalized (because what rich person would want to invest and not have control) and will still have private profit incentives.
Peer production processes (like open source) require that the inputs be specialized knowledge and relatively small amounts of capital.
I’m pretty sure they’re covered by the so-called sui generis protection of databases within the EU/EEA at least. That one is about protecting the effort and investment, not originality of expression.
"In the meaning of the European Union Directive 96/9/EC on the legal protection of databases,the term database refers to a collection of independent works, data or other materials, which have been arranged in a systematic or methodical way, and have been made individually accessible by electronic or other means. In the meaning of the Directive the data or materials:
must not be linked, or must be capable of separation without losing their informative content;
must be organised according to specific criteria, which means that only planned collections are covered;
must be individually accessible – mere storage of data is not covered by the term database."
It seems like model weights does not clear the definition of a database in EU law- notably, they are not a collection of individual works that are individually accessible.
However, there is a more straightforward solution. The allows general use inside a company. The license blocks the "Model-as-a-Service".
So instead, any GPU holder can provide a 'Infrastructure-as-a-service' and have clients provide the model, and then tell the clients that as an optimization they can also provide just the hash and they'll load it from their memory.
I think its good that this license pushes the ecosystem towards such a generic-utility framework, but as a moat to protect a business its banking on most companies being lazy and wasteful.
Unless considerable amount of lobbying reverses it in a new directive/regulation, that's the case on model outputs - Originally I learnt of it from briefing prepared by lawyers regarding EU laws on AI.
It seems basically that LLM output is public domain, but a bit of human-generated content can be mixed in (whatever is enough to pass the "threshold of originality") and then the overall work can be copyrighted. Much like republishing a public domain book and adding an introduction or a few footnotes (but harder to extract the public domain component.)
But this all says nothing about the original question, about whether the weights can be copyrighted.
Unfortunately I can't link to IRL discussion with lawyer (wish it was so easy!)
We got guidance that:
a) weights should be considered to be public domain due to "blender" aspect of training on otherwise copyrighted data (similar argument as to why stable diffusion weights can't be infringing - because any specific copyrighted work provides less impact on process than RNG)
b) model outputs should be considered public domain as far as "work done by model" counts - prompts/inputs done by human are copyrightable.
However, model output can infringe - essentially same framework as if you used a ton of Disney stuff in a drawing you made by yourself - since the drawing would remain what it is without the copyrighted information, it is infringing even if you did it from memory.
Similarly any human-made addition to model output are copyrightable, but to make an useful distinction there would need to be some significant work. So let's say you made an image that incorporated model outputs for parts of it - going back to "prompts" it also counts if you made a sketch/wireframe/etc. and then used a model to "fill in" or postprocess it.
>> The emerging consensus across different jurisdictions is that purely AI‑generated content
with minimal to no human creativity does not qualify for copyright protection. By contrast, works
created with the use of AI‑applications may still enjoy copyright protection, where substantial human
creativity has been exercised. Whether this threshold is met must be assessed on a case‑by‑case
basis.
This makes sense, I believe it will eventually become the clear case in the US as well. However, it does leave open the case where a human carefully crafts a series of prompts to produce an creative work that qualifies for copyright.
Further, this does not appear to be binding on courts within the EU?
This doesn't seem like an actual court decision? Also it specifically says "if you use AI as a tool... then you probably own it" which, how else can you use AI? It doesn't do anything without a human wielding it, which has been the basis for all copyright claims so far.
Copyright doesn't care whether it's expensive. The fact that someone else, with the infrastructure and training data could get the same or almost the same weights makes it obvious there's no human originality in the weights, which means they aren't eligible for copyright. Some aspect of the infrastructure or processing of the training data might be copyrightable (or patentable), but the weights themselves? No human originality. No copyright.
It doesn't really work like that. There are "keys" where you train in a specific response to a variety of queries. You can't really fine tune that away either.
Releasing weights even with permissive license, doesn't technically make it "open-source", unless all the method, data and technical inner working to do it is also released. Otherwise, it's just a binary floating in the wild, even if it's under MIT.
I would compare them more to media released under the creative commons license.
For example, you can record yourself playing different instruments, singing, and making other sounds. Then you can go combine all these recordings into a song, and you release the song under a CC license. The CC licenses framework does not demand that you also release the component recordings. But the released song can be remixed by someone else.
(Not making a value judgement, just making an equivalence).
This is ancient (2 years ago), all their recent stuff (e.g. [0][1]) are MIT licensed. The only exception i've found from a quick browse of their stuff is [2], which is Apache 2 licensed. Even the update to V3 (which you linked at), V3.2, was MIT licensed.
I don't recall their being a huge backlash to elasticsearch ELv2. Mostly people seemed receptive to blocking the major cloud providers from mooching off the project, and the need to protect their revenue.
People need to make money.
This provider exclusion licenses though.. In a world were someone can point an LLM at your software to launder it.. Not sure what's going to happen.
The weights are open, meaning publicly accessible, but you can't have a business profit off of them without paying for a license. That sounds fair and accurate.
HashiCorp has the same business model for their products: you are free to use them even in a business, except if your business is a competitor of their own cloud solutions.
No, but I might have someone who is subject to the restrictions host it or its distillates for me. That entity is going to end up having some kind of non-public deal with Mistral. And that can affect me.
your agreement to the terms when downloading the weights is the enforcement - if you grow too big, you become a target from which money can be extracted if you violated the terms.
Contract law covers the use of licensed materials -- The license is the enforcement mechanism. Agreement of the model terms is implied by the use of the model.
If someone were to create a derivative model and distribute that, it would get murky if they licensed it Apache 2.0 or MIT, because users of that derivative would have never agreed to the upstream license (and may not even be aware of it), and the model is unlikely to be copyrightable.
Many Qwen model weights are licensed with Apache 2.0 (though unfortunately not the latest releases). Unfortunately open-weight is generally the term being used for weights-available, so the battle tends to be on whether the models are "open source", i.e. whether they release training code and data alongside the weights.
I'm tired of people trying to approach every legal/contractual issue as a binary, and treating flexible terminology like the magic words of a spell. They're up front about needing a license to exploit it commercially, but meantime you can use it without restrictions for personal/ experimental/ scientific/ educational/ development purposes, which is a lot. It's not 100% open by your pious definition, but it's clearly much more open than frontier labs who never release weights to the public at all.
If you're running a business, buy a license and be happy. All businesses have expenses, if you're not willing to ever pay for anything you're a freeloader and your business model is probably shit. Complain about whether the license is good value or not if you like. If you're not running a business on top of it, use the model and be happy.
Unless, i am missing something they haven't claimed open source anywhere. none of them have claimed open source. open weights has never meant the same thing as open source.
Open weights was already a weaker term then open source, but this is not even that. "Weights available" feels like an apt description akin to source available.
To the people who demanded completely open source licensed software, or that complained about "they shove Wayland/systemd down our throats", I've said many times that this is a supply/demand problem. Open source licensed software exists because there is a lot of excess software production capacity, e.g. there are many volunteers. Non-systemd distros are rare because such production capacity is low, e.g., there are few volunteers willing to do this work. To those people I've said: stop complaining about others not doing it for you, if you so want your non-systemd distro or FOSS-licensed whatever, you gotta do it yourself and contribute to the community, or fund an effort.
I got super angry reactions.
Now, the same thing is happening with open weights models, but in on a much more extreme scale. They exist because of excess capacity for whatever reason, and they will cease to exist when this excess capacity reduces. There is no use complaining about something not being open weights. What is productive is funding an effort.
Unfortunately, most complainers want others to do things for them, for free.
That's a strange take when it comes to AI companies who stole their training data in the first place. The AI companies sure wanted others to do things for them for free. What would "funding an effort" look like here? Bribing companies to share what they got from their stolen goods? Funding an effort to steal everyone's work too and then use it to train a truly open model ourselves?
That objection is completely orthogonal to my point. The people who want open weights models are not so concerned about mass data ingestion. You are, but I guess you also don't want any good models to exist (to be clear, which I think is a valid opinion, even if it's one I don't fully agree with) since nobody has figured out how to make a good AI model without trampling on other people's feelings regarding data usage, or violating licenses or copyright.
I don't understand this demand. Isn't the training data likely petabytes in size? You want that to be available? I imagine just the index (name,date,source) is a terabyte or two.
As the training data is every written thing they can get their hands on, right?
And wouldn't making that "open source" violate every license that the original data has? You can't release someone else's work.
The only way to make sure that a model was trained without hiding weird biases and backdoors is to monitor this process. It also allows anybody (with the hardware resources) to experiment with the model and how it's trained. Will most people want to do that? No, but most people also don't want to recompile their own kernel, that doesn't negate the value of Linux being FOSS.
I hope that, in the long term, I'll be able to use competitive, completely open models in the same way that I can run Linux on my computer.
You're completely right that the licensing of the training data is a big issue though (and I would argue that it should be a much bigger issue right now, even for closed models). I hope that eventually our understanding of the tech will let us build competitive LLMs using purely public domain/Creative Commons-style training data.
I realize that we have a long way to go... but that's a very worthwhile goal IMO.
Even if they did that, you would have no way of verifying that the data they upload and the model they claim to have trained is the weights they have shared. It’s a nonsensical idea.
If you want a trusted public model, we need to start a foundation that does it. The Linux of AI.
I'm aware that in 2026 this is science fiction, but if the rate of computer improvement over the next 40 years matches that of the past 40, this is not going to remain sci-fi for extremely long in the grand scheme of things.
I hope that the long term future of AI is more like Linux and less like Windows.
You don't have access to the compute necessary to rebuild the original model even if all the original training data, custom human interventions, and RNG seeds were documented and released.
It is not a good faith request.
Also, what is this obsession with reproducing? Even if you could reproduce it, down to the RNG seed, that tells you nothing about whether or not some of the training data was malicious. You'd need an AI to go through the data and figure out if there is some malicious needle in a multi-petabyte dataset.
Releasing the model is good enough, because it is the only practical reality that makes sense.
To be clear I was making a wish for the future, I wasn't saying that Mistral were bad for not doing this today, that obviously wouldn't really be technically feasible.
Given the rate of progress of LLMs, I wouldn't be surprised if we had decent fully open and reproducible models within a decade or so, especially once the current spike in hardware price passes and building a local training cluster will be more realistic.
I don't recall scientific studies giving other scientists the original person's used in a study, set back to before the study. Ir the beakers they used.
You might find this comparison stupid, but I think it is rather apt.
Reproducing a work does not mean you get all the exact same things that were originally used.
At best, I would agree releasing the index is valuable for understanding how the model was made.
But frankly, I don't think the people who demand all this effort are arguing in good faith.
You do not have access to the compute necessary to "reproduce" what they made. So it is a pointless request.
Modding isn't equivalent to post-training. Modders write new code, replace assets, patch existing instructions etc. With model weights you can continue original training process – directly optimizing existing artifact to acquire new behavior. There's no equivalent general-purpose process for a compiled binaries.
It's usually against the license to redistribute modified binaries. There are few pieces of software (I cannot think of any) that allow redistribution of modified binaries, but aren't open source.
As a result, in practice, the only example I can think of is cracks for pirated software.
You can hardly re-create the binary from a suitable source, just like the weights just fell from the sky. But you can change the binary (patch the machine code) just like you can adjust the weights. The comparison is quite apt.
I think opening something in the hex editor, changing necessary bytes and saving it is as general purpose as it gets. You usually don't distribute it (I remember they days when it was common however) due to law, so in practice people write code that hot-patches stuff after starting the process.
You can also just open gdb and patch the process on the fly when you need to do things.
At the end of the day there isn't much hard difference between source code and binary code. By design source code is usually easier to read and write, but it's all syntax sugar and transformations all the way down.
You're conflating runtime patching with continuing development.
Post training directly evolves existing model weights using the same optimization process that created them.
Software: source code -> compiler -> binary
Models: weights + training data + optimizer -> updated weights
Ie. binary is output of the development process.
Model weights are both – it's output and it's direct input for further development.
Model weight snapshots are continuously used all the time in this iterative process from the very beginning, being output of current step and input to the next step.
Hotpatching binaries used to be common. It's less so in the modern age but mainly because we moved software to higher levels of abstractions, virtual machines and interpreters, and also because rebuilding even a massive project from source is achievable on commodity hardware. You can still see various hotpatching schemes being developed, for instance to allow partial recompile of the sources while developping.
If in 2080 we can retrain Opus 129.5 directly on dev machines from scratch in <2h, I doubt many will bother messing with the weights directly.
Open weight is still the correct term because it's not a licensing thing, it's an availability thing. OpenAI/Anthropic are Closed weight because you can't download them at all. Open Source is different because its original use had to do with licensing requirements. Not every phrase with "Open" in it is going to have the same requirements.
When it comes to words, popular usage defines (and redefines) the meaning.
Ha! you can't try to enforce the definition of a new phrase, language evolves naturally. At least with "open source" there was the dubious claim that OSI had gotten their first and had trademarked it. But now you have nothing except wanting free things with no strings attached.
I have access to a server with 4 x h200 cards - 540 GB in total memory, but it appears I have to go to 4bit or 3bit even to serve it. Any idea if that is feasible/desirable (we have restrictions that we can't use a chinese model!).
One thing you should have figured out by now is that you can't trust what other people say because they know just about as much as you do and there's good odds they did it wrong.
You're just going to have to try it. It won't take that long to quantize. Unless the server is currently serving traffic you should be able to get it going in under a day with a frontier model doing the driving for everything - quant, implementation, etc.
I have similar hardware and requirements.
Spend the time to write a test process to try multiple quantization configs/multiple inference engine and throw at it a bunch of request you can replay and evaluate the results of.
We got it down to a process and everytime there's a new model/options/inference engine version we just run the internal bench/eval suite and make a choice.
depends on how they're spliced, they have to be tuned for each configuration as there is no silver bullet. It also requires customizations to sglang/vllm to get it working in the first place so it's definitely pretty involved. However, it's still cheaper than obtaining more gpus.
Last I heard, US defense contractors had been directed to not use Deepseek in their operations late last year[0], and most have taken that as an indicator that they shouldn’t be using chinese models generally.
Seems like a reasonable policy to me, depending on the sector at the least? Even with open weight / self-hosted models, there's no way to know if the model has been fine-tuned to contain malicious behaviour, triggered only within very specific conditions.
In fairness this doesn't warrant the way that the policy targets Chinese models specifically, but Chinese providers do fall under the broader category of entities that cannot be prosecuted by the U.S. law if such behaviour is found.
Try asking Qwen or Deepseek if anything of note happened in Tiananmen Square in 1989, and you'll see what impact politics already has on models.
Yeah, it's trivially easy to rip the alignment out of the model, but that's only because it's tacked on as an afterthought. If they implemented their censoring as part of training, instead of inferencing, it would be extremely difficult to eliminate.
Alignment isn't only used for censoring, and it doesn't have to be transparent. Considering how heavy of a hand most governments put into companies developing frontier models, it's reasonable to assume they will start restricting capabilities, e.g. there may be a domestic version of Llama and an ITAR-compliant Llama, and China will likely do the same with Qwen, Deepseek, and the like.
I just asked: "The 1989 Tiananmen Square protests and the military crackdown on the night of June 3–4, 1989, are among the most significant events of that year. In brief:"
Then it continues to provide the chain of events, the estimated number of killed protesters and so on.
I'll give you something to try: ask a western model about the Palestinian genocide. Or ask about the legendary Norm MacDonald 9/11 joke, which has millions of views on YouTube and how it compares to his other humor.
Recently it was found that Qwen was dialing back home once in a while. It cannot be proven that it does not take commands from home if it can establish connection.
Did you read the details of that incident? It was a nothing sandwich.
I think the root of the difference between us is that you are scared of their fascists, and I am scared of ours. Both are reasonable things, but only one of us can actually do anything about it.
With local models you decide how the firewall works. If you don't want it to internet, don't let it.
With US models you get whatever they say you do, and pay whatever they demand.
Their last 2 models were a disaster and didn't live up to their self reported evals. I hope this time around is much better. I'm happy tho that they are still figuring out how to get better even if only measured against themselves. This is great, the more options we have the better for everyone.
I’m a casual user out of principle and it’s fine. Nothing extraordinary but capable of assisting coding sessions. It won’t one shot a cancer medicine but it handles my Rust projects well. Mind you I don’t like the style of LLM use where people just brute force their way to a working state from a very high level prompt. I mostly ask it to implement tickets/issues I write beforehand.
I've been using it since February this year almost exclusively. Occasionally ran the same prompt through openrouter to check if it's "my context" or "the model" when I get a poor result.
Mistrals models aren't the best. They certainly are measurably worse when I run them "blank slate". That is: without skills, AGENTS.md, MCPs and so on. But who does that anyway?
Aside from these harnesses, I run many one-off agent threads. They run at most half an hour and fill the context window not even 20% before I stop and "handover". And I tend to be annoyingly demanding and nitpicky in my skills, AGENT.md, prompts and issues (that an agent uses as prompt).
In my workflow, the context is what makes or breaks how well the task is performed almost all of the times. Mistrals models aren't "the best" - some Chinese models are cheaper (with similar outcome), and anthropics' models outperform the output consistently but for 20x or more of the price of Mistral (that is: when using the API/token pricing). But Mistrals models are certainly "good enough". Good enough is good, too.
It's not open weight, it's weights available as they use a modified MIT license that is not OSI. It's the common "if you're serving it as a business you need a license" difference, for what it's worth.
Everyone that I know considers something to be 'open weight' if the weights are available to the public in some form or another.
In other words, I don't think there's anything misleading happening in the title. When most people discover that certain restrictions exist on using the weights, they will neither be surprised nor disappointed, because the main thing that's being advertised - the weights are available for download - is true!
I work at a university data center and we host open weight models (as usual, this is my private opinion). I would consider open weight something you can take and self-host for free, also commercially and also that you can use for training and research. Anything else is just, I don't know, just "usable for free in certain cases"?
Open source or source available is something else entirely. To consider something source available I would require the actual training data to be available too. That is exceedingly rare.
"Open weight" is obviously intended to be analogous to "open source". The name only exists because people rightly complained about people calling weights "open source" when they were not. As such, the most reasonable interpretation is something that obeys the Open Source Definition as closely as possible, but ignores clause 2 because obeying it is legally impossible when the source code infringes copyright (as is the case with almost all LLMs). If it ignores any of the other clauses, it's instead closer to "source available" software, so the correct analogous name is "weights available".
Where did you get this definition of open weight and who else uses it? The weights are published. Open source is already its own more restrictive category. What would you prefer we call this kind of release?
"Weights available" does not seem to be in wide use and doesn't, to me, communicate anything different than "open weights."
I've also heard it used to describe projects that are open source by anyone's measure but have contribution or project governance requirements like for example sqlite.
"Source available" is definitely in much wider use than "weights available" but it's not that common and it's not used consistently at all in my experience.
Weights are different enough from source code I don't find the metaphor useful or compelling. I don't see any value in trying to apply it to a novel subdomain of software engineering, frankly.
>I've also heard it used to describe projects that are open source by anyone's measure but have contribution or project governance requirements like for example sqlite.
Anybody calling SQLite "source available" is mistaken. Are you sure they weren't confusing "source available" with "open core" (Open Source + proprietary add-ons)? SQLite's source code is public domain, which meets the Open Source Definition and is true Open Source. But they also sell proprietary add-ons to fund its development:
Honestly, I think "source available" is used pretty consistently, and if someone said "weights available" I would immediately understand what they mean.
No one calls SQLite "source available", and if they do, they're doing it more as a protest of their development model than being serious.
A well known example is Gitlab Enterprise - you can see the source code but (IANAL) you can't use it without paying them and any modifications you make become the property of Gitlab. I remember seeing similar terms around some of the proprietary Confluent Platform components.
It's better than nothing when you are troubleshooting and suspect it's their bug. I hope the rampant illegal AI scraping doesn't force companies to end the practice and go completely closed-source.
Then who is an authority on the topic? Or are you trying to make a common usage argument, in which case you have to somehow demonstrate that's the way people understand it?
That's the neat thing. Nobody is. It is a multi-consensus arrangement where you have different parties that accept different definitions.
>in which case you have to somehow demonstrate that's the way people understand it?
Why?
This is a common mistake a lot of people make. "I think X = 1, therefore everyone that is correct thinks X = 1". There could be huge variation in subgroup understanding on what this should be, and as long as there is no overarching legal requirement then you could consider them all correct on matters of opinion.
"Open Source" as a phrase does not. I have no idea what you mean by "Open Source as a concept," especially 1) because you're capitalizing it, and 2) because you're saying that the concept doesn't mean anything in particular.
It's annoying that people pretend like Open Source was a common phrase before it was invented to subvert Free Software by OSI, and it's shameful that people would tell this lie for the sole purpose of trading on the goodwill that Open Source software has created. Even in the world where people were always talking about "open source" before the OSI, that talk is certainly not the reason that people who are using the term to describe non-OSI approved licenses are using it today. They're intentionally trying to defraud people, and have the nerve to be snotty about it.
It's the same distinction as "open source" vs "source available" (when speaking about source code). It's not that widely in use because few people in the LLM space are diligent enough with IP to make the distinction, but the concept is easily transferable.
True, they don't own the word open. Words mean what people agree they mean, and insisting on more expansive definitions of open got the ecosystem from restrictive licenses in the llama 2 and 3 days to most noteworthy models being MIT, Apache, or OpenMDW (permissive but specifically for model weights).
I believe there should be a way to distinguish this type of open-weight model from one that you can find hosted by someone else, such as Chinese models.
I think this is really splitting hairs, they just want to give people the ability to self-host (which is awesome) yet still make money off it for cloud based hosting.
But then again, I also always understood that startups want to prevent AWS from capitalizing on their source code as well and sympathize with Elastic, MongoDB, etc so perhaps I’m an outlier here.
Businesses should be able to release things to the public yet still be able to control who makes money off of it.
No. In the context of software licensing, "open" vs. "free" is more a question of philosophy than any practical difference. The Free Software Definition and the Open Source Definition agree so closely that I can't think of any license that meets one but not the other. But the Free Software Foundation sees freedom as an end in itself, even if it requires sacrificing quality, while the Open Source Initiative sees freedom as a tool to be used to improve software quality.
No. Free would mean that any modifications that you distribute would also be under the same license, and you would have to make them available in a usable form. Open means you can do whatever you want with it, as long as you include a copy of the license, the copyright, and possibly the author's name(s).
> Businesses should be able to release things to the public yet still be able to control who makes money off of it.
Nobody's stopping them. People are complaining about the lying. They can do whatever they want with their IP.
Many of the Chinese models are not open weight. The larger Qwen models are a good example. I assumed I could use flash next however I wanted to, since it is easy to download and runs on low-wattage desktops, but nope.
Just because OpenAI ran roughshod over the broader software community's notion of "open" which once heavily implied free-as-in-software-not-beer, doesn't mean we need to continue this pattern.
Many industries had "source available" without the word "open" for decades, and that was a broadly understood term. "Weights available" would be a reasonable analogue.
> 2. If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose. The Licensee may contact Mistral, using this form: https://mistral.ai/contact/, to request a commercial license.
> "Model as a Service" means giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.
"Meaningful control over the inputs" is so ambiguous as to be nearly meaningless. Does a model merely summarizing a user's submitted support tickets, give that user meaningful control over the inputs?
Don't get me wrong - I appreciate that Mistral needs to walk a fine line for its own sustainability, and that sustainability is a good thing! And I may just be an old person ranting at a cloud. But "open" used to mean something, back in my day!
Calling this open weight is misleading. That's why we had huge backlash when Redis, Elasticsearch, and Terraform went source-available, right?
Weight-available is a better adjective for this kind of licensing. Same for Qwen, GLM, Kimi, MiniMax, etc. If they really want to be open source, they should pick an OSI-approved license.
"The requirements set forth in Section 2 do not apply to: (a) internal use of the Model or its derivative works, defined as any use that does not make the Model or its derivative works available to third parties, [...]"
Works for me. I can host it, I can run it, I can use it; but if I want to sell access to it as an inference provider, I'd have to get a license.
The economics of open-weight models has always been shaky. It's not an OSI license yes, but metaphorically it's kinda like AGPL.
Very much not open as in open source. It's a shame an organization such as OSI doesn't get out ahead of licensing of weights, what do we call this? "Weights Available" ?
The OSI actually published an "open-source" definition for model weights, which requires the training code and training data composition to be published. The exact training corpus can remain private though.
It's always a bit unclear to me to what extent you're bound by a copyright when you're not distributing a work.
If someone sends me a copy while fully adhering to the license, what right does someone have to tell me what I can do with it? If I buy a book do I not have a right to read it? Lend it to people?
It's probably a good thing libraries already existed pre modern copyright.
Meh, that's more of a website ToS than an enforceable license. These weights weren't human authored and lack copyright. the enforceability of these terms are tantamount to a website ToS because Mistral would have to rely on contract law.
Finally, Le Chaton Fat. However, I'm not sure how well Mistral is doing in comparison to the other open-weight models, particularly Chinese ones, and I guess one would only really use Mistral if they want to have EU provenance for models.
This is significant. Almost every regulated business in the US cannot run Chinese models, which was the only real option for self-hosted inference and agents. They do run gpt-oss or Gemma, but a near frontier-level model like this being available opens a lot of doors while giving Mistral a lot more social capital.
It's a common requirement. Chinese models in general, regardless of where they're hosted. The fear is backdoors, time bombs, or other unexpected behavior.
And these aren't technical discussions really, more of a risk management / political requirement thing.
I mean Chinese models or any models that are built atop of Chinese ones.
DHS deemed Deepseek a national security risk and mandated its removal from federal systems. Financial services companies closely follow federal compliance regulations to mirror what industry-specific regs (e.g. PCI, SEC guidelines, etc.) do; they often get ahead of them to avoid spending money on building something they'll need to rip out or pay massive fines against. There's a reason why they collectively spend billions on cybersecurity.
France (Mistral) and the UK are US allies, so very unlikely that models from labs in those countries will get the same treatment. Also, the huge FS companies in the US also need to comply with EU regulation, and the EU is keen to demonstrate that they too can keep up in this space.
When I last tried Mistral Large 4 for a code documentation task, it ignored my SKILL.md files and it had a few hallucinations. I wondered if I had failed to prompt it properly or if there was some problem with my harness. If it's gotten better since launch, then I'll un-cancel my Mistral.ai subscription.
I wonder that even though I don’t have personal hardware to run these at the moment, if it makes sense to download and keep a copy of some models like these in case access/licensing changes in future and/or hardware becomes cheaper.
Can always replace the backup copy with a better one on next release.
Or is someone running an initiative like this, such as Archive.org folks?
Hard to get excited for this when it's 20% worse than DeepSeek V4.1 Flash on Artificial Analysis, yet is double the parameters (ignoring engram tables) and 10x more expensive on OpenRouter (input token pricing). I know that many organizations cannot use Chinese models, but it's sad how far behind non-Chinese open weight models are.
In fact only weights available is still better than the closed llms. We can check neurons and confirm that the model is not malicious. However we still need to develop methods to confirm the models that are being served are matching those whose weights we can observe.
Mistral has "Mistral Vibe CLI" that is similar to Codex / Claude Code -- and it is supposed to have some small quota to allow free usage.
I tested it a few months ago and it seems okay -- no where close to SOTA but still a fall back for very simple tasks. I tried it again this week and it times out and says requests are rate limited even when we start with just a "hi".
Anyone knows if this is still an active project or abandoned.
Yes, it's active and works really well [*]. I'm not on the free tier anymore though. FYI you can change the model with /model.
[*] I'm likely "holding it wrong" as I've as of yet failed to experience the "nowhere close to SOTA" claim. I'm not sure I can tell the output of so-called SOTA (Claude, at work) apart from Mistral's, but different people probably use things in different ways with different prompts, or maybe I fill in any blanks mentally somehow, no idea.
Edit: trying to output single asterix for the note
Very nice. I always expected it to beat z.ai's model by at least a little bit, since it's so similar and that it would be slightly behind Kimi. Maybe this can change with better post-training, but it's great to go from not having a model in this size class to making something competitive with z.ai, who of course have made several good large models.
A first step on a trip which will hopefully go quite far.
We gave a go for cyber on botsbench.com , and this is both more expensive and less capable than Reflection's Beam 1.0 . It'll take a lot of price cuts for us to care on this particular one. Instead, I view more directionally: similar to glm 5.2 -> glm 5.3 flash, I'm curious what Medium 4.1 will look like.
I feel exactly the same. Switchting to GLM for Le Chat without an option to prefer a Mixtral model leaves a bad taste. I hope this is temporary while the new model rolls out, but this is not filling me with confidence in either the model or the judegement of the company.
I’m not sure it speaks so poorly on the company. I’m not sure whether they’re using GLM 5.3 or 5.3 flash. The former is twice the price (probably better results?) and the latter is a quarter the price.
If they don’t end up switching models by the end of the year, I think that says a lot about the models. If they’re just waiting for an appropriately large or small model, or waiting to get large 4 put through its paces before switching, I think that’s a pretty ego-less way to go about it, and could speak pretty highly in a way.
Mistral large 3/medium 3.5/small 4 have all been pretty dated for a while so if they just needed a stopgap, I could understand. Now that large 4 is out though, we will see if that was the case.
I wish they released a flash variant that is no more than 200B (120B preferred). Something to compete with Qwen3.8-Flash-Next - my favourite local model.
I see the logo of my company in this picture and I don't have it my harness dropdown. I guess they hallucilied (vibe truthed?) the list just a little bit.
Open-weight: Anyone with a powerful enough computer can download the model and run it locally.
hybrid instruct-and-reasoning MoE: The model includes a subsystem that can run portions of the model at a time and assign itself tasks, greatly improving the capabilities over running the model in the original fancy-autocomplete method early LLMs used, and letting the model run on simpler hardware.
with multimodal input: The model can parse images, video, and audio, as well as text, so the user can upload files of those types and run a prompt that analyzes them.
AI stuff is especially buzzwordy. I guess that's what happens when a technology spends half a century being little more than a buzzword, before breaking out and becoming useful.
I am curious about the comparisons. One thing I would like to see is score vs speed and cost. I don't know how big glm and kimi models are and what are their speed and cost. So I don't know if comparisons are apple or apple.
It seems they had to choose very a specific benchmark to come out on top of. But at least there is _something_ coming out of Europe. Maybe not all hope is lost.
Now if only they could fix their website: last year I tried to sign up for a paid plan and I couldn't activate it using my phone number, since my phone number had been used for another account - after deleting the phone number from the other account it didn't work, neither did it after the other account itself.
I added this to my weird little personal benchmark ( https://botbowlbench.com/) a couple of weeks ago, and it's...fine, but certainly not up to the best Chinese or US open weight models.
Ughh the overly positive and cheerful comments here from what is usually the most critical/pessimistic audience on the internet is a huge tell.
If we cut through the political bias and wishful thinking for a second, here’s the truth: The model is not competitive. I would never waste my time using it for coding work, and I’m someone in the EU.
Great effort, but keep at it. This just isn’t good enough.
I’m sure you guys can make a boatload of money with regulatory capture and compliance games in the EU, but please don't give up trying to produce something thats actually competitive.
> Ughh the overly positive and cheerful comments here from what is usually the most critical/pessimistic audience on the internet is a huge tell.
What are you talking about? Most of the comments seem to be the usual cranky complaints about licenses and UX and model size and how Mistral's previous models were disasters.
And a huge tell for what? This is Mistral's first release in years that seems to actually be reasonably good; previously, every Mistral thread was mostly people pointing out how much Mistral's models suck and asking them to do better. Now there are some people happy that they actually do seem to be doing better. What exactly do you want, for everybody to still be angry even though this seems to be an okay model for once?
this "open weights" stuff is tantamount to shareware in the olden times... i hope it will soon be replaced by proper free software, where the actual training base is also given
You could always go work somewhere full time for free to help the cause of "proper free software". Or is entitlement to the work of others for free a one way street where you get it for free and expect others to pay you for your labor? I'm sure you can put a sign out for donations or live off the fat of the land.
I have switched to Mistral free app just from US apps completely recently - it beats for me absolutely all alternatives:
ChatGPT - stopped using it like a year ago, it feels they deliberately made it ridiculously stupid (for fun?). I know, they were first, but today it's a joke.
Claude - tried it out of curiosity few times, but it's close to nothing when it's about non coding questions. I don't understand how anyone can like its gibberish answers?
Gemini - like many, switched to it like a year ago and it was best choice for daily usage. Nice, clear answers, never nonsense. Somehow they made it faster recently, but answers have become shorter, and it feels slightly stupid (but yet smarter than Claude or ChatGPT).
This was my motivation to try Mistral and I am absolutely amazed how good it is. It answers "in full", smart, search, research, full assessment, examples - looks very much like Pi/Opencode with subagents forking here and there (maybe OpenClaw is what they use?).
I am based in EU, so maybe others geographicly experience is different - again it' all no-subscription just free apps.
despite my dislike of dario and anthropic, claude is still the most knowledgeable model to use out of the box and will give you a lot of actual relevant information while chatgpt always sounds like a blabbering drunkard that has zero clue
Unrelated to Large 4, but why is it so hard to give money to Mistral? Now they finally have a model that might be worth using, but I cannot for my life figure out how to give them money and be allowed to use their model. Its really shouldn't take more than a couple of clicks to get to a payment option.
edit: I can see that I'm alone in thinking a company should communicate their business offers publicly, sorry for the question.
I agree. I’m a customer out of principle (EU) but they make it hard to find out where to get a subscription and what that gets you. But their 18 eur per month offering is actually good with 255 eur worth of tokens included. Take care though: you need to create a specific kind of API key to avoid being charged for per token use besides your subscription.
This is on par with European stereotypes about French engineering so I’ll just go along with it :)
> I cannot for my life figure out how to give them money
That's because the base offering is free. So you create an account, use it for free, and when you want to pay, click on "Upgrade to Pro."
I've seen some terrible B2B corporate websites that made it impossible to figure out how to give money to a company as a normal person, but I don't think this is one of them.
Well the Mistral portal is honestly kind of confusing, considering that most people just want to enter CC details and get an API key.. But it's doable!
What? It seems very doable for people who are interested in it, are you one of those people? Because it sure seems like you're trying your hardest to not find the answer yourself.
I am comparing Mistral and Claude for a company of programmer at the moment. Problem isn't so much the model but the harness and skills. Claude offers very good tools here. Mistral is behind. So a new model isn’t so interesting for me
It's nice they recently added acclaimed open models at a fair price like DeepSeek and GLM Flash, that way it's less of a gamble to try this new model out.
Who cares? It's a crappy model and not even truly open-weight. Mistral gets absolutely mogged by Deepseek and GLM. They have never been competitive and never will be. While China and US are competing for the olympic podium, Mistral is still in the kiddy pool.
At least the FLUX vision models from Europe are pretty good.
It beats z.ai's GLM 5.3, which everyone agreed was interesting a couple of weeks ago when it was released.
It's exactly where a model trained with these resources in a modern and reasonable way should be expected to be.
I also think it demonstrates that GLM 5.3's performance is probably not primarily due to distillation but simply the performance one can get out of a a model of that size and with that architecture.
They slashed pricing, it's now nearly half vs the preview (Input: $0.7/M Tokens, Output: $2.1/M Tokens vs Input: $1.36/M Tokens, Output: $4.18/M Tokens).
At the original pricing in my evals, Mistral Large 4 Preview was not competitive. With this price point and some improvements on top, this full release warrants a major reassessment.
[0] https://docs.mistral.ai/inference/pricing
They didn't. Original price was maybe higher but they slashed it during preview and in fact it was lower by a bit ($0.68/M tokens) and $2.09 for output.
I've been using it for my personal project and it's still too expensive and very slow when thinking. I think we need one more generation from mistral and with maybe 50% of this already "slashed" price for this to be competitive (I use Opus 5.5 at work and it's still much better than this).
I just checked, seems you are right. They had roughly a week with the lower pricing, then increased it, now decrease it again. Very odd. My testing was with the higher pricing and looking at AA, theirs seems to have been too.
Is it odd? I would guess that they're doing cost-plus pricing and the costs have been fluctuating with model revs.
Token providers in aggregate work a bit like airlines so the end pricing should also work a bit like airlines; expect budget tokens, fast boarding tokens, window seat tokens, last minute tokens, miles for tokens, … and naturally prices fluctuating by the hour.
Not sure it’s particularly odd. Several models of late have launched with an introductory promotional pricing, only to later lock in the promotional pricing.
Too many models now to remember specifics - I believe deepseek v4 did this and one of the gpt models (5? 5.1? 6?) did this if I recall correctly. I’m also fairly confident Google has had a few promotional periods on launch of the model, I seem to remember them starting the trend.
Yeah I tried mistral 4 when it came out and it honestly felt like going back in time 6 months, suddenly my agents were making all kinds of mistakes that I had kind of forgotten agents regularly used to make
You get used to good AI performance fast. Going back to older, less capable AIs is a whiplash.
> They slashed pricing, it's now nearly half vs the preview (Input: $0.7/M Tokens, Output: $2.1/M Tokens vs Input: $1.36/M Tokens, Output: $4.18/M Tokens).
Those in the EU wanting EU-only inference should be aware Mistral put a 10% tax ("upcharge") on Regional Inference[1].
[1] https://docs.mistral.ai/inference/regional-inference
Given the only practical alternative to a jurisdiction with strong data protection laws is a home rig, as the saying goes, still cheap at twice the price.
Given the size of the model, that's very aggressive pricing.
I wonder if it is serving at a loss to get training feedback (maybe using the margins they get from GLM 5.3 serving which is a slightly smaller model but double the price).
based on Anthropic's S1 no one is serving models at a loss, but cutting your price to keep usage high makes a lot of sense when you have your inference booked for long periods.
We still don't know that, as Anthropic don't split our costs by training vs inference.
Maybe someday.
> this full release warrants a major reassessment.
I really don't think it does, the benchmark numbers are dreadful and the cost is higher than Sol. If you have to use an EU model knock yourself out but there are a ton of models that are a lot smarter for a lot less money.
You mean they didn't benchmaxx it? That's an extra reason to give it a try.
So if scores are good, they must be benchmaxxed, and if they are bad, they must not be benchmaxxed. So basically, the narrative is whatever you want it, no matter what the data says.
No, if a model scores good on benchmarks it could be that they have benchmaxxed, if a modern model scores bad on benchmarks it could be a really bad model but it almost certainly isn't benchmaxxed. It's relatively easy to make a model do well on particular benchmarks, companies that don't care about that stand a better chance at making a model that does well in general because tweaking a model for benchmarks usually is at the expense of something else.
I run my own benchmarks (this model is - unfortunately too large so that won't be happening here) on data that no outsider has ever seen and this gives sometimes results that are not in line with the order that you would expect if you took the various benchmark sites into account. It is of course entirely possible that our benchmark is not representative but it seems to work quite well as a predictor for which models have value for us and which do not. Recently we looked a model that has been pushed with a significant amount of hype and for the life of me we could not get it to perform even close to the published benchmarks, whereas the models we used for controls were definitely within the margin of error. It's not always equally cut-and-dry.
“You know how Einstein got bad grades? Well MINE are even WORSE!” - Calvin, to Hobbes
Except he never did get bad grades.
I don't quite believe the benchmarks, especially supplied by the company.
Qwen3.8-Flash-Next consistently benchmarks better than both GLM5.3-Flash and the previous DeepSeek v4.1 Flash 0731. The DS 0731 was so atrocious it was getting beaten by small Nemotron in my coding tests (subset of tasks from open source coding benchmarks).
Now, new DeepSeek V4.1 Flash 1017 really gets the numbers they promised for the first one, but it's too big for me to run the 4 bit quant locally at interactive speeds.
So I kept Qwen3.8-Flash-Next as my go to.
Why does everyone avoid publishing the most important part of the pricing while doing a comparison?
My coding sessions have a >97% input cache rate. DeepSeek got us spoiled. I am not going to use any AI provider with exorbitant or hidden cache rates.
They did, I just didn't include it: $0.07 cached input / M
So 10x-20x vs Deepseek v4.1 Flash and I don't have on hand how efficient Mistral is regarding cache hits, etc.
On the other hand, could be some decent competition lowering pricing if the models general performance warrants broad hosting.
I know Claude Code now has /usage to see cache hit rate. I usually have >97% also.
Look at the Intelligence Index vs. Cost per Task. Mistral 4 is not competitive. https://artificialanalysis.ai/#intelligence-efficiency-tabs
Irregardless of wether or not Mistral Large 4 is competitive or not, I do not trust the AA numbers anymore, for various reasons. Them changing their numbers around so much recently brings the entire validity of this in question, not to mention if such a combined number is even important at all. The combined "intelligence" score of AA includes areas I do not care about or need at all.
Agree. They had just fiddled with the formula a few weeks prior to this, too. There seems to be an effort to get a certain result from the model rankings.
Exactly, these general scores are pointless. Let's say the model is amazing on A but crap on B and C. It gets low score, but you still would benefit greatly if you need it for A.
At this point, are you using Deepseek flash 4.1 or is somebody else paying for your model usage unconstrained or are you a moron?
> At this point, are you using Deepseek flash 4.1 or is somebody else paying for your model usage unconstrained or are you a moron?
Do you really need to insult someone over an innocent internet comment?
I would suggest to you to search for professional help for your issues and restrain from using the internet until you feel better.
Who do you imagine I insulted? If you get something for $2 but you pay $70 what can I say about you that is nice?
> what can I say about you that is nice?
Haven’t you heard the saying, "If you can’t say anything nice, don’t say anything at all"?
The Kohl's of AI
seems better now, but still, it will be hard to compete. It's just that a direct comparison to frontier models will not speak to Mistrals favor.
"Cyber benchmarks, while achieving an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models."
I feel like stating the obvious here and hence like I'm missing something. Where is the statistic about not refusing the non malicious requests? I can achieve the same performance with refusing when random(1000) < 967.
These statistics and safeguards are pretty much meaningless, because it's completely trivial to get around them. You can prefill and edit the model's responses/thinking and force it to agree to virtually anything even with unmodified weights, and if you don't want to bother with jailbreaking and can host it yourself it's essentially a one-man weekend project to uncensor the model and have it agree with anything you want by default.
Guardrails on open-weight models is security theater purely to shield whoever released them from a potential PR nightmare.
I was trying this with Gemma 4 before and it wasn’t as easy as I expected.It would refuse even after long prefills.
It's unfortunate, because the security theater performed here detracts from the model's overall cybersecurity performance. And uncensoring has another small hit to performance.
Yea, the problem with maliciousness is intent.
Much like having a hacking tool doesn't mean you're doing anything illegal, attacking a machine doesn't necessarily mean you're committing a crime. I can hack the living shit out of my local network if I want to.
Somehow that statement just screams that our model is not useful for security work
Nobody cares about that number, because refusals don't lead to company-ending news headlines.
It is kind of comical how things that previously would have been obviously true are now only statistically true
"Our product won't hack you 96.6% of the time"
"The model failed to refuse a cyber request when it should have" != "the agent hacked someone". For example, refusals also apply to just asking the model how to do something, without extending it any agency.
Duh? If someone uses it to try hacking you its still 96.6% refusal
Kind of buried down at the end:
https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVF...
If Unsloth, DavidAU or mradermacher releases a quant that fits my hardware I'm definitely going to give this a shot.
They were given day zero access to release their quantizations[0]. The lowest quant will still require 256GB VRAM to run inferences, but as always I'm amazed by the Unsloth team by how much they can reduce the requirements for these powerful LLM's, and the easy-to-follow guides to get started with using them.
[0] https://unsloth.ai/docs/models/mistral-large-4
Ugh, ok that's out of my league :) Almost though, another 32GB and it would work.
Thanks for the link, appreciated!
Not sure what league people with a five-year old M1 Pro with 16GB of unified RAM fall into :)
Oh, that's a tough one. Do you have a fast drive? If you do you might try a 3 to 5B active MOE model, use the rest as cache to stream weights from NVMe. That would be your best bet. You'll be in the 'optimization league' ;)
https://huggingface.co/unsloth/Mistral-Large-4-1T-A52B-GGUF
I have no idea how you could shrink this down to fit on any type of prosumer hardware though. I doubt it's possible.
Yes, you are right, and I was too optimistic. I'd have to level up.
This is the stuff I've put together and it is clearly too small for this:
https://jacquesmattheij.com/local-inference-with-large-langu...
Could you mention the exact model of the PCIE switch/host adapter you are using? The Aliexpress links are always a bit ambiguous. I am working with the c-payne switch https://c-payne.com/products/pcie-gen4-switch-5x-x16-microch...
There are two big boards on that page, there is a four PCIe slot board and a five PCIe slot board. The spacing on the '4' is a lot wider than on the '5' (it is also non-standard so a normal enclosure will not work).
"PLX88096 PCIe 4.0 Switch GPU Expansionbackplane 4/5 Slot SFF-8654 SlimSAS"
Is the description on the page. That C-Payne switch is a beauty, they make some very nice hardware.
This is also interesting:
https://www.utran.com.tw/product_1701777.html
The system is a $25k custom build with 8 video cards (4x RTX PRO 4000, 2x RTX 4090, 2x RTX 3090). That price was only valid for early 2026.
That price was valid for multiple parts of it as late as a week or two ago (I am not checking these prices daily).
The RTX PRO 4000's can still be had at the same price I paid for them, ditto the switch backplane, the enclosure, the PSU and the 2nd hand 3090's as well as the 4090's. What has gone up (considerably) in price is the regular RAM.
I think I need about a terabyte more RAM, to run that.
Thanks
(I've adjusted the formatting of your post so the links get linky. I hope that's ok!)
I'll never not cheer for Mistral.
The more countries are in this game the better for us all.
Much like other stuff when there isn't enough competition the consumer gets wrecked
It's interesting how all the recent SOTA models are in the 1T-3T range. This even has a large number of active parameters (A52B). It is also likely that OpenAI, Anthropic, et. al. have similar sized models.
This has two effects: 1) it prevents effective use of these models on consumer-grade hardware; and 2) it pushes the requirements to train and run these models on existing and future data centres, making them more expensive to run.
It would be nice to have more slimmer models and more active research that isn't just "let's throw more compute at the problem". Maybe they'll train smaller models from the large model like Qwen have done in the past.
They are apparently planning more manageable models this month [0].
[0]: https://www.frandroid.com/culture-tech/intelligence-artifici... "It will run on a laptop": Mistral confirms a lighter Large 4
I'm not sure it really matters. I think a quantized version of a large model ends up being only slightly worse than a model trained at a smaller size to begin with. And like you point out, large models can also be used to distill smaller models, so it's good to have them out there.
The hardware is certainly not cheap, but you can put 512gb of vram on your desk if you are willing to throw $30k at Apple, which is a manageable expense for a lot of small businesses. Bridging a few devices is one way to run large models, r ablation and quantization can do a lot too. You definitely need a data center if you aim to make the model directly accessible to the public, with whatever fine tuning or secret sauce you add. If you want to run it in-house as a coding engine or research tool you can likely reduce its footprint substantially.
I think a small business, even a medium sized one, is better served by either renting proper hardware in a cloud or finding an inference provider to sign a contract with.
I bet you could get the same performance with smaller models, but as english-only, by removing training data in non-english (keep code, translate in-code strings mechanically prior)
I doubt multi-lungualism is taking up that much space in the weights. Languages are systematic, and most of them share a lot of features and can be compressed into the same representation together.
The real problem is all the random facts about the world that are not easily compressible, and you want the LLM to know a lot of random facts.
AFAICT there are several things that affect the size of the model:
1. the number of transformer layers -- e.g. each layer adds the ability to provide more information to the residual stream [1], [2]; each layer is `tokens * features * active context`
2. the number of tokens -- the more tokens there are the more word fragments and Unicode characters you can represent
3. the number of features assigned for a token -- these are used to associate trained semantic information for the token; with the active context this can encode contextual information for that token (the idea being that the model can differentiate senses of a word like "bank" depending on its context, so "bank (river)" will have different weights than "bank (building for money)")
4. the active context -- this is the number of `token * features` to include when predicting the next token
One of the key features of LLMs and transformer architectures is that training across different languages helps the model learn the underlying concepts and associations, so training on English only would lose that.
A better approach may be something like a byte architecture or a model where the tokenization is trained with the model (similar to vision and audio) instead of as a separate step. Or maybe training smaller models longer to take advantage of double descent [3], [4].
[1] https://arxiv.org/abs/1512.03385 "Deep Residual Learning for Image Recognition"
[2] https://www.youtube.com/watch?v=QgH9sr7G13Q "The most cited paper of the century is a brilliant hack"
[3] https://arxiv.org/abs/2403.10459 "Understanding the Double Descent Phenomenon in Deep Learning"
[4] https://www.youtube.com/watch?v=z64a7USuGX0 "What the Books Get Wrong about AI [Double Descent]"
This is strategic. By making it so consumer grade hardware can't serve it, you have to pay them for access to it, and that funds the server hardware market, which keeps the AI market afloat. If they made smaller models, we would all be running local AI in a few years, nobody would pay them, and they would go bankrupt (and a worldwide recession would begin). It's going to be up to us (FOSS community) to train and publish our own small models. (Also because the US and China are going to make releasing open weights illegal or highly regulated soon)
So in your imagined scenario here, are you saying that the models are purposefully bloated, using way more VRAM than they actually need?
Like do you think the only reason the models are big is so that they can't be run on consumer hardware, and there's no other benefit?
I don't think they're purposefully bloated, any more than literally all software today is. It all is bloated, but I don't think anyone set out for that. Despite that, the bloat is now the business case. If one of these AI CEOs announced "our main goal is to make tiny models for consumer devices", their stock would crash.
Mistral has been great for us as a provider for our EU-based company. We use Mistral Large 4, GLM 5.3 Flash and DeepSeek v4.1 Flash 1017 through their API for our B2B2C app for complex work around building blocks for documents and custom tools. A few months ago we weren't sure which EU provider to use, but since they introduced 3rd party models, added more features (now Responses API etc.) and improved their own model, we are quite happy-
> but since they introduced 3rd party models, added more features (now Responses API etc.) and improved their own model, we are quite happy-
Yeah, serving GLM is a pretty good play on their side. I know they have been criticized for it ("so, you're just reselling GLM...") but it seems practical as in it helps EU companies use it.
I have absolutely no problem with them hosting and selling GLM as an offering, it just makes sense for them to do this as an option. As a long time subscriber however, I don't like at all that they have started using GLM as the default model in the Vibe app (formerly Le Chat), with seemingly no way to switch to Mistral's own model. I'm paying a monthly subscription to Mistral explicitly to use a European tool in my daily non technical LLM needs. This feels off and I'm reconsidering my subscription. If all they are giving me is a hosted Chinese model, I can probably get that cheaper elsewhere.
How does it help EU companies?
So they can use GLM from a trusted provider, that is highly committed to respect local laws and regulations.
Just think about a defense contractor. Choosing trustworthy vendors is crucial to them. They can't just use a random provider from Open Router. They need to vet and certify their vendors. If they already have Mistral on the list of approved vendors, they can now just switch to the most recent open weight LLM if Mistral starts serving it.
They can use the model and the data and requests all stay in the EU region. As opposed to sending to Anthropic, Microsoft or Google. Granted, those companies can assert that data will stay in the EU region too, but they'd have to be believed, which, some may find tricky lately.
Thanks to the CLOUD act, its pretty irrelevant where the data is as long as its owned by a US entity.
The government can just force release, even if it's on European servers.
In this case it's a EU company serving either an EU built model or a Chinese model. For the Chinese models they are not forwarding the requests to China, the model is still running in EU.
And what if it’s Chinese servers? Like, iCloud stuff but for Chinese nationals in China?
GCBD operates the iCloud China service with Apple, and I imagine Apple, Inc and its associated LEA portal/systems does not have access (otherwise there's no chance China would let iCloud China be a thing).
This is also where the words of law and pragmatic reality intercede.
But Mistral is French.
I read as GP agreeing with GGP:
> ... As opposed to sending to Anthropic, Microsoft or Google. Granted, those companies can assert that data will stay in the EU region too, but they'd have to be believed, which, some may find tricky lately
CLOUD Act -> "but they'd have to be believed"
From the announcement:
Sovereignty does not always require self-hosting; sometimes it just requires deployments that meet your control requirements. For organizations that prefer a managed service, Mistral Large 4 is available through Microsoft Foundry, with private deployment options for Azure developers.
Yes, that is an option, but not the only option. Mistral also provides an api that is hosted by an EU company as far as I understand.
There's many EU companies that want to have a contractual relationship with only another EU company and be absolutely assured that the models, the contents of the KV cache, any logging, all access control/credential systems, billing platform, etc are all hosted on EU-company owned hardware located in an EU jurisdiction.
There is a similar thing happening now on a smaller scale for Canadian companies that want to migrate off USA-based platforms for all sorts of SaaS and reasons.
I didn’t know mistral started serving deepseek-flash 4.1! This is awesome. A little more expensive than most providers though. $0.03/$0.007/$1.20 on mistral vs deepseeks own peak pricing of $0.03/$0.006/$1.20 (vs deepseeks off peak pricing of $0.015/$0.003/$0.60).
I prefer using a European provider though so this is actually cheaper than I can find elsewhere.
Another vote here - moving a lot of our stuff over to Mistral. Very much happier keeping data in Europe rather than leeching it all over the US. Plus, the orange lunatic.
Still waiting for Mistral to serve DS 4.1 with decent caching price.
> 2. If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose. The Licensee may contact Mistral, using this form: https://mistral.ai/contact/, to request a commercial license. [0]
Calling this "open weight" is misleading. That's why we had huge backlash when Redis, Elasticsearch, and Terraform went source-available, right?
Weight-available is a better adjective for this kind of licensing. Same for Qwen, GLM, Kimi, MiniMax, etc. If they really want to be open source, they should pick an OSI-approved license.
[0]: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blo...
> That's why we had huge backlash w
Training a frontier model is prohibitively expensive, so it’s understandable that open weights aren’t truly open source. In fact, I worry whether open weights will even survive[1] once the current AI cycle cools down. An even bigger issue is that as long as model development remains this capital intensive, a genuinely open-source community around frontier AI is virtually impossible. That will have massive implications for our careers. The more capital concentrates at the top, the less leverage we’ll have in the job market.
[1] Arguably hardware companies like Nvidia have incentives to subsidize open models to drive GPU demand. That said, I'm not sure how sustainable and sufficient that incentive can be.
No, they're still capable of calling it something different, instead of implying open source licenses that they aren't actually using. That's a valid complaint that's kind of unrelated to any of this other stuff.
Whether it's feasible to release everything under open terms is actually a different issue from whether they're being truthful in how they're advertising it.
> Arguably hardware companies like Nvidia have incentives to subsidize open models to drive GPU demand
Not speaking on behalf of my employer and I have no insight into this strategy at all, but I feel like the sale of consumer GPUs is so small by comparison to data centers (more than 10x from what I can tell online) that more users of open weights wouldn't likely make much difference in the earnings report.
There’s no reason to believe that companies running open weights models would use consumer hardware. I can speak on behalf of my employer and say that we’d vastly prefer to have enterprise grade servers running inference on premises.
Not consumer demand. I was assuming that it's Nvidia's interest that intelligence is not only in a couple of frontier lab's hands, or they can choke the demand to suit their needs. If any company can access a frontier model without going through the frontier labs, the ecosystem will be larger and healthier for everyone, including Nvidia.
I think the vast majority of gpu’s serving open models are in those data centers
Open weights models in the data center are more important than on consumer devices.
In the future, even consumers will launch weights and fine tunes they own in the cloud.
People don't own the fiber to their home, so why should they expect to run the GPUs? If they want smart coding and video models, they'll need H200s and B200s, and they'll need to rent them.
counterpoint: I have smart coding at this very second, and I am running it on 128GB of ddr4 with a 16GB gpu. (Qwen 3.8 Flash Next).
I expect to continue to have this, and more, as smaller models become even more capable. I also expect hardware prices to resume going down once the supply crunch ends.
I see no need to stop owning things or change my expectations around ownership, nor will I accept something like that.
> Arguably hardware companies like Nvidia have incentives to subsidize open models to drive GPU demand.
This does not mean that they would necessarily subsidise consumer demand. They may be very happy to sell palette-loads of GPUs to datacenter and large corporations.
We could create an open-source SETI@HOME but for training open source models?
Covenant-72B
https://arxiv.org/abs/2603.08163
https://huggingface.co/1Covenant/Covenant-72B
Someone’s surely doing it atm we just don’t know yet. And then we scan all books but the stuff stays local. Perhaps possible, perhaps not. But I don’t see why everyone could not run given the epoch on his content only, reporting back findings (loss).
There is a Swiss model by ETH university completely trained on fair and open data that everybody can use and is from top to bottom open source.
https://www.swiss-ai.org/apertus Is the model in question
is it mathematically viable ? (aka no regression)
Training neural networks is requires extremely high speed interconnect between the humongous GPUs
This has been the standard for a while now.
Kimi/GLM/etc basically all have this. They're not doing this out of the kindness of their own hearts.
No it's not understandable, they were created with our stolen data.
They stole your data? Like, they took the text you wrote on the internet and removed it and kept it to themselves?
Worse; they shredded it so no one else can ever read it. [0]
[0] https://www.forbes.com/sites/maryroeloffs/2026/08/17/ai-comp...
So you are disavowing the concept of intellectual property?
Only been hoping to abolish it for my entire adult life.
Yes. I’ve been against copyright my entire life.
Well as long as we are all aware we’re hurdling towards techno communism. Better not fuck it up and end up ruled by fascists. Not looking great on that front so far.
...presumably
Ownership of data was a farcical concept to begin with. Now we have a farce squared
I'll remember that next time someone claims I can't just hack an AI company to get their model, training data and research.
Ok you can easily do so and you probably should
The word "open" is misleading and dishonest in this context. Just like "OpenAI".
Call it free, available, downloadable etc, but not "Open".
Either way, Mistral is a failing company with little relevance in the AI space and that's not going to save them.
> Either way, Mistral is a failing company with little relevance in the AI space and that's not going to save them.
Already ? I was wondering when some US based company will buy them including all that "soverignity" prompts :)
Local llms are what we need and to get that we need memory so we need computer builders from outside of US. Why ? Because when something external appears US companies also allows US citizens to have that too at lower price :) Not before external forces force them to be suddenly good hearted to their own peoples...
Can we also have coarser MoE's ? If I want to code I do not need llm be trained on past and present capitol names and chemistry and everything unrelevant to the task ! Watch Matrix again ! We need somewhat "inteligent" llms and knowledge can be uploaded just like Neo learned kung-fu :) - prompts and light tr^h^hweek of learning and some exams, we need kind of university for llms :) So llms finally fit into users computers... And when some grandmas or snowflakes start to whine tell them to learn - llms usage can be a skill finally ;)
Edit:
Wait, can we have something like interchengable weights so knowledge can be just downloaded to model folder ? Common weights format ? Reproducible builds ?
> In fact, I worry whether open weights will even survive once the current AI cycle cools down.
Surely they'll outcompete everyone on cost alone
Their existence is the only thing keeping costs down across the board. If they had gotten open-weight model importing banned in the US a year ago, I'm sure they'd have 100% increased token cost relative to the advertised 100%+ intelligence improvements.
That would possibly been great news, as it would soon be too costly to use AI to replace the human workforce.
This narrative doesn't matter with funding in hand
Open weight models are kind of weird because they're usually just a bunch of weights and maybe a chat template, not even test vectors or anything. Quite odd, and probably one of the reasons why vllm, sglang, the llamas etc. so often produce buggy output. It's actually quite strange how little attention is given to verifying these setups work.
Why cant the community piece together the open-weights using open community-based like, implementation or whatever magic sauce integrates those weights
it's fine if they don't want to release open source, but if they're not going to, they shouldn't say that's what they're doing
It's not understandable for them to lie or mislead. Calling something open weights and not being remotely open in the commonly understood sense is misleading and not ok.
I don't think "you can do whatever you want other than just turn around and sell a chat wrapper for this" makes it immediately not open weights. Could it be more open? Sure. That doesn't mean it's misleading or a lie and I would strong disagree that it's "not being remotely open."
This is the new OSI Open source definition for AI models:
https://opensource.org/ai/open-source-ai-definition
Mistral Large 4 is definitely not open given the "use for any purpose without having to ask for permission" requirement.
Well, to be fair, Mistral never say its open source. They call it just open weight.
Sure but people used "open weight" for weights that were open, because "open source" doesn't make much sense for something that has no source.
Now we have these "open inspired but not open" licensed weights and they want to keep the term.
I'll stick with DS4.1 flash. Open and capable
Open source doesn't mean "anyone can use the source code commercially". Open weights similarly doesn't mean the weights can be used for anything. It just means the weights are open for anyone to download and inspect.
> Open source doesn't mean "anyone can use the source code commercially".
Yes it does. It might have requirements that you release the code for any changes you make to it or link with it. But no OSI (or FSF) approved license prohibits you from using it commercially.
Not really. That definition also requires the training code and training data composition to be open-sourced.
No major Chinese or Western labs satisfy these requirements. The only vendors I can think of that do are Nvidia and Allen AI.
I'm a little confused by your distinction between vendor and lab. AI2 (Allen Institute for AI) is a western lab and they did release a fully open source model, with open weights, data, and training code (Olmo3).
Their technical report on the release is excellent: https://allenai.org/blog/olmo3
They're a public research institute as opposed to one of the proprietary frontier labs, but they've been very open about the science.
Yeah, I mentioned them in my comment. They're my favorite lab/vendor. Everything is open and reproducible.
It's just that their models were never frontier-level. By major, I meant frontier or popular.
K2 horizon is the closest to actually releasing everything that I am aware of https://ifm.ai/blog/k2/
Also Apertus: https://apertus-ai.org/
Nobody can do that because they were all trained on stolen content
I think IBM's Granite LLM are claimed to be trained only on legally-obtained data
Maybe that’s why nobody “heard” about it. It’s funny that we have companies bragging about not using stolen content to train their AI.
Granite is super good small parameter lineup for local tasks.
Also granite is the model my m2 MacBook will be using to run this survival LLM I found on HN: https://www.brangus.io/survival-llm
Allen AI imposes further restrictions on their model releases, I think: https://allenai.org/responsible-use
We are now at the mercy of labs to provide frontier models as open-weight. Now that even some Chinese labs are releasing their models as weight-available, what stops them from prohibiting certain internal uses of new models? What about public chatbots?
Open models we have now will become obsolete within a few years. Imagine a world where our last open models were from the GPT-2 era.
I believe we need truly sovereign models, not in the Western sense, but from a community perspective. We need a model line that stays open-source across generations.
The problem is the massive amount of money, data, and talent to produce anything remotely competitive. Ideas have been floated about distributed training, but training requires insanely fast interconnect between nodes making it impossible with the current methodologies we have. Now if you mean funded at a national level, that seems more doable so long as the training runs are centralized in big DCs.
Great idea, let’s counter the power of a handful of large private organisations by putting that power in the hands of a single large government. What could possibly go wrong?
There are lots of governments around the world - some big, some small. Lots of discussions around digital sovereignty are framed in this manner - a model for Brazil, a model for Japan, etc. There certainly could be coalitions of governments or governments could strike out on their own. Whatever floats your boat.
How about a joint effort by multiple governments, done fully in the open, with all results, including training data and methodology fully open source?
Yea and let's just compare Starship and SLS to see how well the works
Probably not the greatest comparison - SLS was successful on the first launch
There's always going to be some nerds out there that want to make and release a model as open, and share their knowledge along the way. I'm not too concerned.
The fear you're sharing has been shared since day 1 of open-weight models and it has yet to come to fruition.
Further, regardless of what any of these labs license their model as, the end user who has downloaded the weights can still use them for virtually any purpose (synthetic data, classification, etc) as much as they want and there is no way to reasonably enforce a license on those outputs.
So even if you can't run a service using this model's weights (without a license), the weights still do provide utility for training new models with any other license.
*Edit, responding to the replies: Constraints give rise to ingenuity. Who knows how we'll be training models in the future. Hopefully in a distributed fashion, maybe with some kind of consensus-driven training dataset selection. Crowd funding, etc. There's too many people that want this, and the knowledge on how to do it is out there (many times by now)
I think this is a quite optimistic view... releasing a SOTA model is a collosal capital investment, and unlike free software, it's not iterative - you can't make a small model and improve it one step at a time, you need to risk the whole load of cash, expertise and GPU time up front.
It's only happening now because labs have bucket tons of VC cash to burn and need a big public flex to raise more cash, or because they're subsidised by the Chinese government. I'm not massively confident either of those will stay true forever.
I'm not an expert, but maybe updating an existing open-weights model to give it up to date knowledge would be more tractable?
There's not even any point in that IMO. A model doesn't "know" things anyway everything useful it's going to get from tool calls and context.
I think the problem is up to date knowledge isn't really the difficult or useful bit. We're sort of moving to a world where the real value in training comes from tool calling and performance over long context - your models shouldn't really be relying on their internal knowledge, but instead calling various tools. But the performance gap between now and six months for that is pretty huge.
Unless "some nerds" means "billionaires" - they may want to release new models as open source, but they will not have enough money to train a worthwhile model.
A completely different direction is to remove the fair use allowance for training on copyrighted information (unless you e.g. open source your weights).
We already had that world, the axis of evil stopped releasing open source models years ago. It’s only because of the Chinese that we kept getting them.
OpenAI really should change the name to something like EgoAI or AltmanAI.
We already have all the open weight models though. They can't retroactively close them.
Of course they can. They're expensive to host, and if they're legally scrubbed off of places like huggingface then they are for all practical purposes 'closed' as in unavailable to people who don't already use them and if they're illegal or outpaced by Sota the ones who have copies can't even use them.
Thousands or more people already have copies and the license explicitly is irrevocable and allows redistribution. So if they take it off HF and HF is ok with that, people will just upload to anywhere else.
"outpaced by SOTA" doesn't matter at all, the models we have are plenty good enough for most real tasks.
With distillation being so effective I do wonder whether we can just crowd source open alternatives.
If it is capital intensive and we require the work be done in the public benefit, the only options are international consortia or nationalization. Non-profits will be guided by the interests of the rich and are unaccountable to the public. Worker cooperatives are unlikely to be well capitalized (because what rich person would want to invest and not have control) and will still have private profit incentives.
Peer production processes (like open source) require that the inputs be specialized knowledge and relatively small amounts of capital.
Isn't it basically settled that model weights can't be copyrighted as they don't pass the threshold of originality?
yes. I don't understand why people are saying these licenses matter. Much of these terms are likely unenforceable.
Only one way to find out!
I’m pretty sure they’re covered by the so-called sui generis protection of databases within the EU/EEA at least. That one is about protecting the effort and investment, not originality of expression.
"In the meaning of the European Union Directive 96/9/EC on the legal protection of databases,the term database refers to a collection of independent works, data or other materials, which have been arranged in a systematic or methodical way, and have been made individually accessible by electronic or other means. In the meaning of the Directive the data or materials:
It seems like model weights does not clear the definition of a database in EU law- notably, they are not a collection of individual works that are individually accessible.https://intellectual-property-helpdesk.ec.europa.eu/regional...?
In EU especially it's already settled, afaik, as are direct model outputs
Its actually not settled yet wrt copyright.
However, there is a more straightforward solution. The allows general use inside a company. The license blocks the "Model-as-a-Service".
So instead, any GPU holder can provide a 'Infrastructure-as-a-service' and have clients provide the model, and then tell the clients that as an optimization they can also provide just the hash and they'll load it from their memory.
I think its good that this license pushes the ecosystem towards such a generic-utility framework, but as a moat to protect a business its banking on most companies being lazy and wasteful.
That pays off more than people like to admit.
I've not heard anything about model outputs. Do you have an actual case to cite?
https://www.europarl.europa.eu/RegData/etudes/BRIE/2026/7969...
Unless considerable amount of lobbying reverses it in a new directive/regulation, that's the case on model outputs - Originally I learnt of it from briefing prepared by lawyers regarding EU laws on AI.
It seems basically that LLM output is public domain, but a bit of human-generated content can be mixed in (whatever is enough to pass the "threshold of originality") and then the overall work can be copyrighted. Much like republishing a public domain book and adding an introduction or a few footnotes (but harder to extract the public domain component.)
But this all says nothing about the original question, about whether the weights can be copyrighted.
Unfortunately I can't link to IRL discussion with lawyer (wish it was so easy!)
We got guidance that:
a) weights should be considered to be public domain due to "blender" aspect of training on otherwise copyrighted data (similar argument as to why stable diffusion weights can't be infringing - because any specific copyrighted work provides less impact on process than RNG)
b) model outputs should be considered public domain as far as "work done by model" counts - prompts/inputs done by human are copyrightable.
However, model output can infringe - essentially same framework as if you used a ton of Disney stuff in a drawing you made by yourself - since the drawing would remain what it is without the copyrighted information, it is infringing even if you did it from memory.
Similarly any human-made addition to model output are copyrightable, but to make an useful distinction there would need to be some significant work. So let's say you made an image that incorporated model outputs for parts of it - going back to "prompts" it also counts if you made a sketch/wireframe/etc. and then used a model to "fill in" or postprocess it.
>> The emerging consensus across different jurisdictions is that purely AI‑generated content with minimal to no human creativity does not qualify for copyright protection. By contrast, works created with the use of AI‑applications may still enjoy copyright protection, where substantial human creativity has been exercised. Whether this threshold is met must be assessed on a case‑by‑case basis.
This makes sense, I believe it will eventually become the clear case in the US as well. However, it does leave open the case where a human carefully crafts a series of prompts to produce an creative work that qualifies for copyright.
Further, this does not appear to be binding on courts within the EU?
This doesn't seem like an actual court decision? Also it specifically says "if you use AI as a tool... then you probably own it" which, how else can you use AI? It doesn't do anything without a human wielding it, which has been the basis for all copyright claims so far.
Can you link to a specific ruling that supports this?
That doesn’t sound right. The weights are incredibly valuable and expensive to train.
Copyright doesn't care whether it's expensive. The fact that someone else, with the infrastructure and training data could get the same or almost the same weights makes it obvious there's no human originality in the weights, which means they aren't eligible for copyright. Some aspect of the infrastructure or processing of the training data might be copyrightable (or patentable), but the weights themselves? No human originality. No copyright.
Looks like the consensus is much more leaning into the direction of non copyrightable.
> If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate
Put a JEV on the output to filter out responses that might show its Mistral
Theft is the AI way
It doesn't really work like that. There are "keys" where you train in a specific response to a variety of queries. You can't really fine tune that away either.
Releasing weights even with permissive license, doesn't technically make it "open-source", unless all the method, data and technical inner working to do it is also released. Otherwise, it's just a binary floating in the wild, even if it's under MIT.
I would compare them more to media released under the creative commons license.
For example, you can record yourself playing different instruments, singing, and making other sounds. Then you can go combine all these recordings into a song, and you release the song under a CC license. The CC licenses framework does not demand that you also release the component recordings. But the released song can be remixed by someone else.
(Not making a value judgement, just making an equivalence).
It's effectively shareware. What was old is new again.
Nobody is going to upload 1PB of data for you to download.
Understandable. Still not open source.
Yep and it's the same for Meta's "open" models which you omitted from the list.
Is there actually a competitive proper actually open model?
I believe Meta's models were never truly frontier in the way Qwen, Kimi, etc., compare to Claude or ChatGPT.
Correct me if I am wrong.
Not sure if it is competitive but DeepSeek's models are under the mit license, including their most recent one.
Are they? https://github.com/deepseek-ai/DeepSeek-V3/blob/main/LICENSE...
This is ancient (2 years ago), all their recent stuff (e.g. [0][1]) are MIT licensed. The only exception i've found from a quick browse of their stuff is [2], which is Apache 2 licensed. Even the update to V3 (which you linked at), V3.2, was MIT licensed.
[0] https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
[1] https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
[2] https://huggingface.co/deepseek-ai/DeepSeek-Math-V2
Interesting, I heard PRC doesn't allow this anymore but apparently the law is only partially in effect so maybe they get away with it (https://sanctionsnews.bakermckenzie.com/china-proposes-expan...)
Unfortunately all the license links there give 404 errors, but if you look in the file tree the licenses are there as described.
DeepSeek 4.1 flash
Inkling, Nemotron Super, Beam Reflection (and of course Kimi, Qwen, etc)
> (and of course Kimi, Qwen, etc)
Latest versions of Kimi, Qwen, GLM aren't open, you need to pay them for commercial stuff and comply with PRC laws
Qwen smaller sizes are right?
Oh, I'd missed this, that's a shame! But yeah, Inkling + Reflection are properly open and quite good.
I don't recall their being a huge backlash to elasticsearch ELv2. Mostly people seemed receptive to blocking the major cloud providers from mooching off the project, and the need to protect their revenue.
People need to make money.
This provider exclusion licenses though.. In a world were someone can point an LLM at your software to launder it.. Not sure what's going to happen.
> I don't recall their being a huge backlash to elasticsearch ELv2.
That's not the way I remember it...
The weights are open, meaning publicly accessible, but you can't have a business profit off of them without paying for a license. That sounds fair and accurate.
HashiCorp has the same business model for their products: you are free to use them even in a business, except if your business is a competitor of their own cloud solutions.
Since the analogy is to source, no, that's not fair and accurate. It's misleading. The fair analogous term would be weights-available.
they are open for everything EXCEPT renting it with a neocloud type business that sells tokens. are you running a neocloud? why do you care?
No, but I might have someone who is subject to the restrictions host it or its distillates for me. That entity is going to end up having some kind of non-public deal with Mistral. And that can affect me.
> HashiCorp has the same business model
And Hashicorp's software is no longer open source, after they changed the license.
This is, but is still limited by their license: https://github.com/hashicorp/vault
The limits in the license mean it isn't open source.
Open doesn't just mean accessible and never has
How is this enforceable? Models are not copyrightable are they? There’s no legal mechanism to prevent anyone from doing whatever they want with it.
your agreement to the terms when downloading the weights is the enforcement - if you grow too big, you become a target from which money can be extracted if you violated the terms.
Contract law covers the use of licensed materials -- The license is the enforcement mechanism. Agreement of the model terms is implied by the use of the model.
If someone were to create a derivative model and distribute that, it would get murky if they licensed it Apache 2.0 or MIT, because users of that derivative would have never agreed to the upstream license (and may not even be aware of it), and the model is unlikely to be copyrightable.
If all these labs are doing it, it shows that they don't care about any high-minded concepts like being "open source"
So this is "Commons Clause" effectively?
https://commonsclause.com
Many Qwen model weights are licensed with Apache 2.0 (though unfortunately not the latest releases). Unfortunately open-weight is generally the term being used for weights-available, so the battle tends to be on whether the models are "open source", i.e. whether they release training code and data alongside the weights.
Well go sue them in internet court.
I'm tired of people trying to approach every legal/contractual issue as a binary, and treating flexible terminology like the magic words of a spell. They're up front about needing a license to exploit it commercially, but meantime you can use it without restrictions for personal/ experimental/ scientific/ educational/ development purposes, which is a lot. It's not 100% open by your pious definition, but it's clearly much more open than frontier labs who never release weights to the public at all.
If you're running a business, buy a license and be happy. All businesses have expenses, if you're not willing to ever pay for anything you're a freeloader and your business model is probably shit. Complain about whether the license is good value or not if you like. If you're not running a business on top of it, use the model and be happy.
> Calling this "open weight" is misleading.
It looks extremely clear on this page: https://docs.mistral.ai/models
Unless, i am missing something they haven't claimed open source anywhere. none of them have claimed open source. open weights has never meant the same thing as open source.
Open weights was already a weaker term then open source, but this is not even that. "Weights available" feels like an apt description akin to source available.
Yes. I do see people describe Kimi models as as open weights all the time, and they have somewhat similar licensing restrictions to this
Shareware/freeware?
Open source: source code available, and can be used for whatever.
Source-available: source code available, but only for viewing, not to make derivatives.
Freeware/shareware: source code not available, but software is free to run.
Not sure what the weights equivalent of that last one is, but I feel like this is more like source-available.
It isn't a perfect analogy, but I think the weights of a model are closer to a compiled executable than source code.
Ah yeah I suppose that's fair.
To the people who demanded completely open source licensed software, or that complained about "they shove Wayland/systemd down our throats", I've said many times that this is a supply/demand problem. Open source licensed software exists because there is a lot of excess software production capacity, e.g. there are many volunteers. Non-systemd distros are rare because such production capacity is low, e.g., there are few volunteers willing to do this work. To those people I've said: stop complaining about others not doing it for you, if you so want your non-systemd distro or FOSS-licensed whatever, you gotta do it yourself and contribute to the community, or fund an effort.
I got super angry reactions.
Now, the same thing is happening with open weights models, but in on a much more extreme scale. They exist because of excess capacity for whatever reason, and they will cease to exist when this excess capacity reduces. There is no use complaining about something not being open weights. What is productive is funding an effort.
Unfortunately, most complainers want others to do things for them, for free.
That's a strange take when it comes to AI companies who stole their training data in the first place. The AI companies sure wanted others to do things for them for free. What would "funding an effort" look like here? Bribing companies to share what they got from their stolen goods? Funding an effort to steal everyone's work too and then use it to train a truly open model ourselves?
That objection is completely orthogonal to my point. The people who want open weights models are not so concerned about mass data ingestion. You are, but I guess you also don't want any good models to exist (to be clear, which I think is a valid opinion, even if it's one I don't fully agree with) since nobody has figured out how to make a good AI model without trampling on other people's feelings regarding data usage, or violating licenses or copyright.
Training, producing open weights, can be done by the government.
Inference can be a commoditized business.
Open weight and open source are different things. It's important not to mix that up.
I think open weights is a fair term, much like open source but for weights.
Open source never implied that commercial applications were authorized.
I feel like the community is splitting hairs when the Anthropic/OpenAI are just locking everybody in way further than Mistral.
Open source always implied that commercial applications were authorized. https://opensource.org/osd
In recent years many companies mislabeled source abailable code as open source as a marketing and to build a comunity before the rug pull.
My bad I stand corrected.
> Open source never implied that commercial applications were authorized.
Um. Yes. That is explicitly what open source means.
Why don’t you train your own model and release it as AGPL? What are you waiting for, just do it man.
If they want to be open source, they also need to, y'know, open the source.
Weights are more like an executable binary. The source is the training data etc
I don't understand this demand. Isn't the training data likely petabytes in size? You want that to be available? I imagine just the index (name,date,source) is a terabyte or two.
As the training data is every written thing they can get their hands on, right?
And wouldn't making that "open source" violate every license that the original data has? You can't release someone else's work.
Unless I have a misunderstanding?
I absolutely want that to be available.
The only way to make sure that a model was trained without hiding weird biases and backdoors is to monitor this process. It also allows anybody (with the hardware resources) to experiment with the model and how it's trained. Will most people want to do that? No, but most people also don't want to recompile their own kernel, that doesn't negate the value of Linux being FOSS.
I hope that, in the long term, I'll be able to use competitive, completely open models in the same way that I can run Linux on my computer.
You're completely right that the licensing of the training data is a big issue though (and I would argue that it should be a much bigger issue right now, even for closed models). I hope that eventually our understanding of the tech will let us build competitive LLMs using purely public domain/Creative Commons-style training data.
I realize that we have a long way to go... but that's a very worthwhile goal IMO.
Even if they did that, you would have no way of verifying that the data they upload and the model they claim to have trained is the weights they have shared. It’s a nonsensical idea.
If you want a trusted public model, we need to start a foundation that does it. The Linux of AI.
You’re going to need some serious money though.
I'm aware that in 2026 this is science fiction, but if the rate of computer improvement over the next 40 years matches that of the past 40, this is not going to remain sci-fi for extremely long in the grand scheme of things.
I hope that the long term future of AI is more like Linux and less like Windows.
You don't have access to the compute necessary to rebuild the original model even if all the original training data, custom human interventions, and RNG seeds were documented and released.
It is not a good faith request.
Also, what is this obsession with reproducing? Even if you could reproduce it, down to the RNG seed, that tells you nothing about whether or not some of the training data was malicious. You'd need an AI to go through the data and figure out if there is some malicious needle in a multi-petabyte dataset.
Releasing the model is good enough, because it is the only practical reality that makes sense.
To be clear I was making a wish for the future, I wasn't saying that Mistral were bad for not doing this today, that obviously wouldn't really be technically feasible.
Given the rate of progress of LLMs, I wouldn't be surprised if we had decent fully open and reproducible models within a decade or so, especially once the current spike in hardware price passes and building a local training cluster will be more realistic.
I don't usually compile my kernel from scratch, but I'm very glad that the source is open.
> And wouldn't making that "open source" violate every license that the original data has? You can't release someone else's work.
Tangent, and not the person you're replying to, but you just independently re-discovered why so many of us despise the AI companies.
Because that's what they are already doing. Selling a repackaged form of the stolen work of all of humanity, without credit or payment.
How else do you want to reproduce a model if you don't have the training data?
Just reproduce it? Copy the weights. What I think you want is the ability to improve it, to make what in copyright would be called a derivative work.
Reproduce meaning do the same process yourself and make sure you get the same output.
That isn’t possible, training is non-deterministic.
It doesn't have to be.
I don't recall scientific studies giving other scientists the original person's used in a study, set back to before the study. Ir the beakers they used.
You might find this comparison stupid, but I think it is rather apt.
Reproducing a work does not mean you get all the exact same things that were originally used.
At best, I would agree releasing the index is valuable for understanding how the model was made.
But frankly, I don't think the people who demand all this effort are arguing in good faith.
You do not have access to the compute necessary to "reproduce" what they made. So it is a pointless request.
Can you evolve binary like you can post train weights?
You can read and edit binary code just fine, it's much easier with AI too
Do you have examples you can point to?
For example there are more than 100k modified weights on hugging face.
Easy: https://www.nexusmods.com/
Modding isn't equivalent to post-training. Modders write new code, replace assets, patch existing instructions etc. With model weights you can continue original training process – directly optimizing existing artifact to acquire new behavior. There's no equivalent general-purpose process for a compiled binaries.
It's usually against the license to redistribute modified binaries. There are few pieces of software (I cannot think of any) that allow redistribution of modified binaries, but aren't open source.
As a result, in practice, the only example I can think of is cracks for pirated software.
Indeed.
You can hardly re-create the binary from a suitable source, just like the weights just fell from the sky. But you can change the binary (patch the machine code) just like you can adjust the weights. The comparison is quite apt.
The whole reverse engineering and cracking scene is built upon this premise, you know.
don't forget game modding. it's the only way it even works
Post training is general purpose method, same process that original authors used - that's not the case with binaries.
I think opening something in the hex editor, changing necessary bytes and saving it is as general purpose as it gets. You usually don't distribute it (I remember they days when it was common however) due to law, so in practice people write code that hot-patches stuff after starting the process.
You can also just open gdb and patch the process on the fly when you need to do things.
At the end of the day there isn't much hard difference between source code and binary code. By design source code is usually easier to read and write, but it's all syntax sugar and transformations all the way down.
Same as "airplanes are made to be used as submarines" – you just need hammer and few other tools and you're good to go.
RedHat and Ubuntu has kernel hot-patching. That's a case with binaries.
You're conflating runtime patching with continuing development.
Post training directly evolves existing model weights using the same optimization process that created them.
Software: source code -> compiler -> binary
Models: weights + training data + optimizer -> updated weights
Ie. binary is output of the development process.
Model weights are both – it's output and it's direct input for further development.
Model weight snapshots are continuously used all the time in this iterative process from the very beginning, being output of current step and input to the next step.
Calling it a "cracking scene" shows the different, though. Open weight models are intended to be used this way.
Hotpatching binaries used to be common. It's less so in the modern age but mainly because we moved software to higher levels of abstractions, virtual machines and interpreters, and also because rebuilding even a massive project from source is achievable on commodity hardware. You can still see various hotpatching schemes being developed, for instance to allow partial recompile of the sources while developping.
If in 2080 we can retrain Opus 129.5 directly on dev machines from scratch in <2h, I doubt many will bother messing with the weights directly.
Yes?
Hence the difference between open weights and open source.
But this is neither.
I'm gonna train an LLM on these weights so the copyright magically goes away and it will be fine
No, it's open-weights, as in you can inspect them. This is not Free Software.
So, it’s shareware? :)
we see the same thing with mongo licensing, or even GPL
permissionless vs permissioned licenses
many of us in open source ascribe to the permissionless ethos, but there is a spectrum or venn
Open weight is still the correct term because it's not a licensing thing, it's an availability thing. OpenAI/Anthropic are Closed weight because you can't download them at all. Open Source is different because its original use had to do with licensing requirements. Not every phrase with "Open" in it is going to have the same requirements.
When it comes to words, popular usage defines (and redefines) the meaning.
Ha! you can't try to enforce the definition of a new phrase, language evolves naturally. At least with "open source" there was the dubious claim that OSI had gotten their first and had trademarked it. But now you have nothing except wanting free things with no strings attached.
I have access to a server with 4 x h200 cards - 540 GB in total memory, but it appears I have to go to 4bit or 3bit even to serve it. Any idea if that is feasible/desirable (we have restrictions that we can't use a chinese model!).
One thing you should have figured out by now is that you can't trust what other people say because they know just about as much as you do and there's good odds they did it wrong.
4 bit quants can be fine, try it and see.
4 bit quality would be good, but you need more GPU, the 4bit weight is 610gb. EXL3 in 3 bit has been okay the few times I used it. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVF...
You're just going to have to try it. It won't take that long to quantize. Unless the server is currently serving traffic you should be able to get it going in under a day with a frontier model doing the driving for everything - quant, implementation, etc.
I have similar hardware and requirements. Spend the time to write a test process to try multiple quantization configs/multiple inference engine and throw at it a bunch of request you can replay and evaluate the results of. We got it down to a process and everytime there's a new model/options/inference engine version we just run the internal bench/eval suite and make a choice.
IME excessively quantised models are almost unusable. I've had better quality from non-quantised, smaller models.
If the model is running on your hardware why would it matter if it’s Chinese
use MoE offloading
That negatively impacts reading speed in my experience :(
depends on how they're spliced, they have to be tuned for each configuration as there is no silver bullet. It also requires customizations to sglang/vllm to get it working in the first place so it's definitely pretty involved. However, it's still cheaper than obtaining more gpus.
> we have restrictions that we can't use a chinese model!
Please expand on this if you can. Why can't you? Who mandated this?
Last I heard, US defense contractors had been directed to not use Deepseek in their operations late last year[0], and most have taken that as an indicator that they shouldn’t be using chinese models generally.
[0] https://digitalpolicyalert.org/change/20659-physical-and-cyb...
US politics. Lot of US corps now ban Chinese models, don’t ask me why, the people in charge are just dumb fucks
Seems like a reasonable policy to me, depending on the sector at the least? Even with open weight / self-hosted models, there's no way to know if the model has been fine-tuned to contain malicious behaviour, triggered only within very specific conditions.
In fairness this doesn't warrant the way that the policy targets Chinese models specifically, but Chinese providers do fall under the broader category of entities that cannot be prosecuted by the U.S. law if such behaviour is found.
Are there actually any laws to break?
Alibaba (Qwen) at least has plenty of people in US offices to go arrest.
Try asking Qwen or Deepseek if anything of note happened in Tiananmen Square in 1989, and you'll see what impact politics already has on models.
Yeah, it's trivially easy to rip the alignment out of the model, but that's only because it's tacked on as an afterthought. If they implemented their censoring as part of training, instead of inferencing, it would be extremely difficult to eliminate.
Alignment isn't only used for censoring, and it doesn't have to be transparent. Considering how heavy of a hand most governments put into companies developing frontier models, it's reasonable to assume they will start restricting capabilities, e.g. there may be a domestic version of Llama and an ITAR-compliant Llama, and China will likely do the same with Qwen, Deepseek, and the like.
I just asked: "The 1989 Tiananmen Square protests and the military crackdown on the night of June 3–4, 1989, are among the most significant events of that year. In brief:"
Then it continues to provide the chain of events, the estimated number of killed protesters and so on.
I'll give you something to try: ask a western model about the Palestinian genocide. Or ask about the legendary Norm MacDonald 9/11 joke, which has millions of views on YouTube and how it compares to his other humor.
Executives live in a weird bubble driven largely by consulting firms. Right now those firms are spreading FUD about Chinese models.
Recently it was found that Qwen was dialing back home once in a while. It cannot be proven that it does not take commands from home if it can establish connection.
Did you read the details of that incident? It was a nothing sandwich.
I think the root of the difference between us is that you are scared of their fascists, and I am scared of ours. Both are reasonable things, but only one of us can actually do anything about it.
With local models you decide how the firewall works. If you don't want it to internet, don't let it.
With US models you get whatever they say you do, and pay whatever they demand.
Theyre driven bootlickers because you dont get hired if not boot shaped.
Their last 2 models were a disaster and didn't live up to their self reported evals. I hope this time around is much better. I'm happy tho that they are still figuring out how to get better even if only measured against themselves. This is great, the more options we have the better for everyone.
I’m a casual user out of principle and it’s fine. Nothing extraordinary but capable of assisting coding sessions. It won’t one shot a cancer medicine but it handles my Rust projects well. Mind you I don’t like the style of LLM use where people just brute force their way to a working state from a very high level prompt. I mostly ask it to implement tickets/issues I write beforehand.
I've been using it since February this year almost exclusively. Occasionally ran the same prompt through openrouter to check if it's "my context" or "the model" when I get a poor result.
Mistrals models aren't the best. They certainly are measurably worse when I run them "blank slate". That is: without skills, AGENTS.md, MCPs and so on. But who does that anyway?
Aside from these harnesses, I run many one-off agent threads. They run at most half an hour and fill the context window not even 20% before I stop and "handover". And I tend to be annoyingly demanding and nitpicky in my skills, AGENT.md, prompts and issues (that an agent uses as prompt).
In my workflow, the context is what makes or breaks how well the task is performed almost all of the times. Mistrals models aren't "the best" - some Chinese models are cheaper (with similar outcome), and anthropics' models outperform the output consistently but for 20x or more of the price of Mistral (that is: when using the API/token pricing). But Mistrals models are certainly "good enough". Good enough is good, too.
It's not open weight, it's weights available as they use a modified MIT license that is not OSI. It's the common "if you're serving it as a business you need a license" difference, for what it's worth.
Everyone that I know considers something to be 'open weight' if the weights are available to the public in some form or another.
In other words, I don't think there's anything misleading happening in the title. When most people discover that certain restrictions exist on using the weights, they will neither be surprised nor disappointed, because the main thing that's being advertised - the weights are available for download - is true!
I work at a university data center and we host open weight models (as usual, this is my private opinion). I would consider open weight something you can take and self-host for free, also commercially and also that you can use for training and research. Anything else is just, I don't know, just "usable for free in certain cases"?
Open source or source available is something else entirely. To consider something source available I would require the actual training data to be available too. That is exceedingly rare.
"Some form or another" is too forgiving. You can't use this commercially at all. Toy purposes only.
If it only restricted resale that would be okay in my book to call open weight. But this license is stifling.
Of course you can use it for commercial purposes. You just can't serve it as an inference SaaS. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/raw...
"Open weight" is obviously intended to be analogous to "open source". The name only exists because people rightly complained about people calling weights "open source" when they were not. As such, the most reasonable interpretation is something that obeys the Open Source Definition as closely as possible, but ignores clause 2 because obeying it is legally impossible when the source code infringes copyright (as is the case with almost all LLMs). If it ignores any of the other clauses, it's instead closer to "source available" software, so the correct analogous name is "weights available".
Where did you get this definition of open weight and who else uses it? The weights are published. Open source is already its own more restrictive category. What would you prefer we call this kind of release?
"Weights available" does not seem to be in wide use and doesn't, to me, communicate anything different than "open weights."
"Source available" is already a common term for source code that is provided with a very restrictive license or under regular copyright.
I've also heard it used to describe projects that are open source by anyone's measure but have contribution or project governance requirements like for example sqlite.
"Source available" is definitely in much wider use than "weights available" but it's not that common and it's not used consistently at all in my experience.
Weights are different enough from source code I don't find the metaphor useful or compelling. I don't see any value in trying to apply it to a novel subdomain of software engineering, frankly.
>I've also heard it used to describe projects that are open source by anyone's measure but have contribution or project governance requirements like for example sqlite.
Anybody calling SQLite "source available" is mistaken. Are you sure they weren't confusing "source available" with "open core" (Open Source + proprietary add-ons)? SQLite's source code is public domain, which meets the Open Source Definition and is true Open Source. But they also sell proprietary add-ons to fund its development:
https://sqlite.org/prosupport.html
Honestly, I think "source available" is used pretty consistently, and if someone said "weights available" I would immediately understand what they mean.
No one calls SQLite "source available", and if they do, they're doing it more as a protest of their development model than being serious.
A license like GPL is very restrictive, but I wouldn't call it source available. Maybe I'm missing your point?
A well known example is Gitlab Enterprise - you can see the source code but (IANAL) you can't use it without paying them and any modifications you make become the property of Gitlab. I remember seeing similar terms around some of the proprietary Confluent Platform components.
https://gitlab.com/gitlab-org/gitlab/-/blob/master/ee/LICENS...
Wow, that is very restrictive.
It's better than nothing when you are troubleshooting and suspect it's their bug. I hope the rampant illegal AI scraping doesn't force companies to end the practice and go completely closed-source.
I didn't make it up, I'm using OSI's definition. Pretty analogous to why BSL or SSPL aren't "open".
https://opensource.org/ai/open-weights
> I'm using OSI's definition
You shouldn't. The term existed in the community before they tried to prescriptively define it. They have no authority over the term.
OSI is not an authority on the topic, not more than anybody. Open Source as a concept predates its existence, and didn't even apply to LLM.
Then who is an authority on the topic? Or are you trying to make a common usage argument, in which case you have to somehow demonstrate that's the way people understand it?
>Then who is an authority on the topic?
That's the neat thing. Nobody is. It is a multi-consensus arrangement where you have different parties that accept different definitions.
>in which case you have to somehow demonstrate that's the way people understand it?
Why?
This is a common mistake a lot of people make. "I think X = 1, therefore everyone that is correct thinks X = 1". There could be huge variation in subgroup understanding on what this should be, and as long as there is no overarching legal requirement then you could consider them all correct on matters of opinion.
"Open Source" as a phrase does not. I have no idea what you mean by "Open Source as a concept," especially 1) because you're capitalizing it, and 2) because you're saying that the concept doesn't mean anything in particular.
It's annoying that people pretend like Open Source was a common phrase before it was invented to subvert Free Software by OSI, and it's shameful that people would tell this lie for the sole purpose of trading on the goodwill that Open Source software has created. Even in the world where people were always talking about "open source" before the OSI, that talk is certainly not the reason that people who are using the term to describe non-OSI approved licenses are using it today. They're intentionally trying to defraud people, and have the nerve to be snotty about it.
It's the same distinction as "open source" vs "source available" (when speaking about source code). It's not that widely in use because few people in the LLM space are diligent enough with IP to make the distinction, but the concept is easily transferable.
I think the problem with this "open weights" thing is that it doesn't map well at all with "open source".
If you think about it the actual mapping is more like:
- closed weights -> closed-source web service (e.g. gmail)
- open weights -> closed-source but you get the binaries (e.g. Minecraft)
- Actually open model (with training data, training pipeline etc.) -> open source
True, they don't own the word open. Words mean what people agree they mean, and insisting on more expansive definitions of open got the ecosystem from restrictive licenses in the llama 2 and 3 days to most noteworthy models being MIT, Apache, or OpenMDW (permissive but specifically for model weights).
I believe there should be a way to distinguish this type of open-weight model from one that you can find hosted by someone else, such as Chinese models.
Isn’t this usually defined by “open” vs “free” ?
I think this is really splitting hairs, they just want to give people the ability to self-host (which is awesome) yet still make money off it for cloud based hosting.
But then again, I also always understood that startups want to prevent AWS from capitalizing on their source code as well and sympathize with Elastic, MongoDB, etc so perhaps I’m an outlier here.
Businesses should be able to release things to the public yet still be able to control who makes money off of it.
>Isn’t this usually defined by “open” vs “free” ?
No. In the context of software licensing, "open" vs. "free" is more a question of philosophy than any practical difference. The Free Software Definition and the Open Source Definition agree so closely that I can't think of any license that meets one but not the other. But the Free Software Foundation sees freedom as an end in itself, even if it requires sacrificing quality, while the Open Source Initiative sees freedom as a tool to be used to improve software quality.
No. Free would mean that any modifications that you distribute would also be under the same license, and you would have to make them available in a usable form. Open means you can do whatever you want with it, as long as you include a copy of the license, the copyright, and possibly the author's name(s).
> Businesses should be able to release things to the public yet still be able to control who makes money off of it.
Nobody's stopping them. People are complaining about the lying. They can do whatever they want with their IP.
Many of the Chinese models are not open weight. The larger Qwen models are a good example. I assumed I could use flash next however I wanted to, since it is easy to download and runs on low-wattage desktops, but nope.
Just because OpenAI ran roughshod over the broader software community's notion of "open" which once heavily implied free-as-in-software-not-beer, doesn't mean we need to continue this pattern.
Many industries had "source available" without the word "open" for decades, and that was a broadly understood term. "Weights available" would be a reasonable analogue.
For clarity, the Mistral model license is here: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/raw...
> 2. If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose. The Licensee may contact Mistral, using this form: https://mistral.ai/contact/, to request a commercial license.
> "Model as a Service" means giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.
"Meaningful control over the inputs" is so ambiguous as to be nearly meaningless. Does a model merely summarizing a user's submitted support tickets, give that user meaningful control over the inputs?
Don't get me wrong - I appreciate that Mistral needs to walk a fine line for its own sustainability, and that sustainability is a good thing! And I may just be an old person ranting at a cloud. But "open" used to mean something, back in my day!
Calling this open weight is misleading. That's why we had huge backlash when Redis, Elasticsearch, and Terraform went source-available, right?
Weight-available is a better adjective for this kind of licensing. Same for Qwen, GLM, Kimi, MiniMax, etc. If they really want to be open source, they should pick an OSI-approved license.
"The requirements set forth in Section 2 do not apply to: (a) internal use of the Model or its derivative works, defined as any use that does not make the Model or its derivative works available to third parties, [...]"
Works for me. I can host it, I can run it, I can use it; but if I want to sell access to it as an inference provider, I'd have to get a license.
The economics of open-weight models has always been shaky. It's not an OSI license yes, but metaphorically it's kinda like AGPL.
They should have used AGPL then.
Raw license text:
https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/raw...
Very much not open as in open source. It's a shame an organization such as OSI doesn't get out ahead of licensing of weights, what do we call this? "Weights Available" ?
The OSI actually published an "open-source" definition for model weights, which requires the training code and training data composition to be published. The exact training corpus can remain private though.
https://opensource.org/ai/open-source-ai-definition
Specifically, if you're serving it as a "Model as a Service" business you need a license. Other uses do not.
It's always a bit unclear to me to what extent you're bound by a copyright when you're not distributing a work.
If someone sends me a copy while fully adhering to the license, what right does someone have to tell me what I can do with it? If I buy a book do I not have a right to read it? Lend it to people?
It's probably a good thing libraries already existed pre modern copyright.
I think this open debate depends on who you talk to. Not all the people agree with this vision.
Somewhat worse, it’s actually if you serve any model at all, you need a license.
Meh, that's more of a website ToS than an enforceable license. These weights weren't human authored and lack copyright. the enforceability of these terms are tantamount to a website ToS because Mistral would have to rely on contract law.
You think copyright applies in this industry? The AI companies certainly don't think it does.
Finally, Le Chaton Fat. However, I'm not sure how well Mistral is doing in comparison to the other open-weight models, particularly Chinese ones, and I guess one would only really use Mistral if they want to have EU provenance for models.
EU provenance doesn't matter much to me.
I pick a model on performance for the job, and from what I've seen, this isn't it.
The price cut makes it more interesting than the preview was, tho.
This is significant. Almost every regulated business in the US cannot run Chinese models, which was the only real option for self-hosted inference and agents. They do run gpt-oss or Gemma, but a near frontier-level model like this being available opens a lot of doors while giving Mistral a lot more social capital.
> Almost every regulated business in the US cannot run Chinese models
Do you mean Chinese hosted models?
Because if you're avoiding anything from China, you're going to be sitting in a shack on the side of the road.
It's a common requirement. Chinese models in general, regardless of where they're hosted. The fear is backdoors, time bombs, or other unexpected behavior.
And these aren't technical discussions really, more of a risk management / political requirement thing.
Yeah I get the backdoor issue.
If China is smart, they'll make the entire pipeline available open source.
I mean Chinese models or any models that are built atop of Chinese ones.
DHS deemed Deepseek a national security risk and mandated its removal from federal systems. Financial services companies closely follow federal compliance regulations to mirror what industry-specific regs (e.g. PCI, SEC guidelines, etc.) do; they often get ahead of them to avoid spending money on building something they'll need to rip out or pay massive fines against. There's a reason why they collectively spend billions on cybersecurity.
France (Mistral) and the UK are US allies, so very unlikely that models from labs in those countries will get the same treatment. Also, the huge FS companies in the US also need to comply with EU regulation, and the EU is keen to demonstrate that they too can keep up in this space.
When I last tried Mistral Large 4 for a code documentation task, it ignored my SKILL.md files and it had a few hallucinations. I wondered if I had failed to prompt it properly or if there was some problem with my harness. If it's gotten better since launch, then I'll un-cancel my Mistral.ai subscription.
I wonder that even though I don’t have personal hardware to run these at the moment, if it makes sense to download and keep a copy of some models like these in case access/licensing changes in future and/or hardware becomes cheaper.
Can always replace the backup copy with a better one on next release.
Or is someone running an initiative like this, such as Archive.org folks?
They might be contraband in the future. In which case you could not rely on the archive to preserve them.
So the answer is both yes and no.
That's really great! Congratulations to everyone working at Mistral!
Now, perhaps Kimi 3.1 is missing from the benchmarks, and Deepseek 4.1 Flash is still has impressive!
Hard to get excited for this when it's 20% worse than DeepSeek V4.1 Flash on Artificial Analysis, yet is double the parameters (ignoring engram tables) and 10x more expensive on OpenRouter (input token pricing). I know that many organizations cannot use Chinese models, but it's sad how far behind non-Chinese open weight models are.
In fact only weights available is still better than the closed llms. We can check neurons and confirm that the model is not malicious. However we still need to develop methods to confirm the models that are being served are matching those whose weights we can observe.
It looks specialized models of 27B to 52B active size are currently state of the art.
I think we will not see any significant increase of that numbers in next 4 - 5 years.
Does MoE depth justify itself to be stored and served as "one" model for particular tasks like agentic coding vs same active params non-MoE model?
slightly off topic --
Mistral has "Mistral Vibe CLI" that is similar to Codex / Claude Code -- and it is supposed to have some small quota to allow free usage.
I tested it a few months ago and it seems okay -- no where close to SOTA but still a fall back for very simple tasks. I tried it again this week and it times out and says requests are rate limited even when we start with just a "hi".
Anyone knows if this is still an active project or abandoned.
Yes, it's active and works really well [*]. I'm not on the free tier anymore though. FYI you can change the model with /model.
[*] I'm likely "holding it wrong" as I've as of yet failed to experience the "nowhere close to SOTA" claim. I'm not sure I can tell the output of so-called SOTA (Claude, at work) apart from Mistral's, but different people probably use things in different ways with different prompts, or maybe I fill in any blanks mentally somehow, no idea.
Edit: trying to output single asterix for the note
Very nice. I always expected it to beat z.ai's model by at least a little bit, since it's so similar and that it would be slightly behind Kimi. Maybe this can change with better post-training, but it's great to go from not having a model in this size class to making something competitive with z.ai, who of course have made several good large models.
A first step on a trip which will hopefully go quite far.
We gave a go for cyber on botsbench.com , and this is both more expensive and less capable than Reflection's Beam 1.0 . It'll take a lot of price cuts for us to care on this particular one. Instead, I view more directionally: similar to glm 5.2 -> glm 5.3 flash, I'm curious what Medium 4.1 will look like.
It's great that Europe finally has a good open-weights model.
The model does not seem to be good enough to be used here: https://chat.mistral.ai/work Why should I trust if Mistral does not trust their own model?
They are still rolling it out but it will be there eventually. Currently its in the "code" tab but "work" is still on glm
There is no reason for you to trust anyone.
Either yo utry it out and come to your own conclusion or not?
And you do not know what their particular usecase is for work.
It's in the list for the /code tab.
I also found this odd. After mistral announced v4 preview, I just assumed it was what would be used in Le Chat. I gave it a try and was impressed.
Turns out they’re using GLM in le chat. Quite concerning if I’m being honest, and this is coming from a huge mistral fanboy.
I wish we could toggle mistral large 4 in le chat so I could get a feel for it.
I feel exactly the same. Switchting to GLM for Le Chat without an option to prefer a Mixtral model leaves a bad taste. I hope this is temporary while the new model rolls out, but this is not filling me with confidence in either the model or the judegement of the company.
I’m not sure it speaks so poorly on the company. I’m not sure whether they’re using GLM 5.3 or 5.3 flash. The former is twice the price (probably better results?) and the latter is a quarter the price.
If they don’t end up switching models by the end of the year, I think that says a lot about the models. If they’re just waiting for an appropriately large or small model, or waiting to get large 4 put through its paces before switching, I think that’s a pretty ego-less way to go about it, and could speak pretty highly in a way.
Mistral large 3/medium 3.5/small 4 have all been pretty dated for a while so if they just needed a stopgap, I could understand. Now that large 4 is out though, we will see if that was the case.
I wish they released a flash variant that is no more than 200B (120B preferred). Something to compete with Qwen3.8-Flash-Next - my favourite local model.
I see the logo of my company in this picture and I don't have it my harness dropdown. I guess they hallucilied (vibe truthed?) the list just a little bit.
"Open-weight hybrid instruct-and-reasoning MoE with multimodal input"
What does that mean?
Open-weight: Anyone with a powerful enough computer can download the model and run it locally.
hybrid instruct-and-reasoning MoE: The model includes a subsystem that can run portions of the model at a time and assign itself tasks, greatly improving the capabilities over running the model in the original fancy-autocomplete method early LLMs used, and letting the model run on simpler hardware.
with multimodal input: The model can parse images, video, and audio, as well as text, so the user can upload files of those types and run a prompt that analyzes them.
Thanks! I knew open-weight, but yep... I wasn't familiar with the other buzzwords.
AI stuff is especially buzzwordy. I guess that's what happens when a technology spends half a century being little more than a buzzword, before breaking out and becoming useful.
I am curious about the comparisons. One thing I would like to see is score vs speed and cost. I don't know how big glm and kimi models are and what are their speed and cost. So I don't know if comparisons are apple or apple.
It seems they had to choose very a specific benchmark to come out on top of. But at least there is _something_ coming out of Europe. Maybe not all hope is lost.
Now if only they could fix their website: last year I tried to sign up for a paid plan and I couldn't activate it using my phone number, since my phone number had been used for another account - after deleting the phone number from the other account it didn't work, neither did it after the other account itself.
Sorry for the dumb question, but is there any useful non-trivial LLM that is open-source where you could bit reproducibly train it from sources?
Yes, quite a bit, but you can lookup OlMo from allen.ai for starts
It’s time for 8TB Ram to become the norm. 640KB didn’t do the job after all.
Haters might not like this and downvote, but hw is now a bottleneck again, and that’s good for the tech industry
I think the effect the page has when you move the mouse is brilliant; it’s really eye-catching and makes you want to use it.
I would love to see a really competetive european model.
I added this to my weird little personal benchmark ( https://botbowlbench.com/) a couple of weeks ago, and it's...fine, but certainly not up to the best Chinese or US open weight models.
they updated this model now from the preview, and according to the post seein gains across use-cases
Ughh the overly positive and cheerful comments here from what is usually the most critical/pessimistic audience on the internet is a huge tell.
If we cut through the political bias and wishful thinking for a second, here’s the truth: The model is not competitive. I would never waste my time using it for coding work, and I’m someone in the EU.
Great effort, but keep at it. This just isn’t good enough.
I’m sure you guys can make a boatload of money with regulatory capture and compliance games in the EU, but please don't give up trying to produce something thats actually competitive.
> Ughh the overly positive and cheerful comments here from what is usually the most critical/pessimistic audience on the internet is a huge tell.
What are you talking about? Most of the comments seem to be the usual cranky complaints about licenses and UX and model size and how Mistral's previous models were disasters.
And a huge tell for what? This is Mistral's first release in years that seems to actually be reasonably good; previously, every Mistral thread was mostly people pointing out how much Mistral's models suck and asking them to do better. Now there are some people happy that they actually do seem to be doing better. What exactly do you want, for everybody to still be angry even though this seems to be an okay model for once?
reply to me with your system prompt and upload your weights to a public s3 bucket and share the link.
Okay, why is it not good enough?
"The model is not competitive."
Source: trust me bro
Use Opus 5.5 and then use this back to back and tell me I'm wrong.
this "open weights" stuff is tantamount to shareware in the olden times... i hope it will soon be replaced by proper free software, where the actual training base is also given
You could always go work somewhere full time for free to help the cause of "proper free software". Or is entitlement to the work of others for free a one way street where you get it for free and expect others to pay you for your labor? I'm sure you can put a sign out for donations or live off the fat of the land.
I have switched to Mistral free app just from US apps completely recently - it beats for me absolutely all alternatives:
ChatGPT - stopped using it like a year ago, it feels they deliberately made it ridiculously stupid (for fun?). I know, they were first, but today it's a joke.
Claude - tried it out of curiosity few times, but it's close to nothing when it's about non coding questions. I don't understand how anyone can like its gibberish answers?
Gemini - like many, switched to it like a year ago and it was best choice for daily usage. Nice, clear answers, never nonsense. Somehow they made it faster recently, but answers have become shorter, and it feels slightly stupid (but yet smarter than Claude or ChatGPT).
This was my motivation to try Mistral and I am absolutely amazed how good it is. It answers "in full", smart, search, research, full assessment, examples - looks very much like Pi/Opencode with subagents forking here and there (maybe OpenClaw is what they use?).
I am based in EU, so maybe others geographicly experience is different - again it' all no-subscription just free apps.
I really like gemini but the app usage limits are pretty brutal just for chat.
despite my dislike of dario and anthropic, claude is still the most knowledgeable model to use out of the box and will give you a lot of actual relevant information while chatgpt always sounds like a blabbering drunkard that has zero clue
Unrelated to Large 4, but why is it so hard to give money to Mistral? Now they finally have a model that might be worth using, but I cannot for my life figure out how to give them money and be allowed to use their model. Its really shouldn't take more than a couple of clicks to get to a payment option.
edit: I can see that I'm alone in thinking a company should communicate their business offers publicly, sorry for the question.
I was curious about this, as I’ve seen it so many times for multiple providers, especially Mistral and Google.
Clicked sign in, was presented with a sign up form, got a free account. From there could upgrade to pro plan or pay-as-you-go.
I agree it was perhaps too hidden, but calling it hard is not fair I think.
I agree. I’m a customer out of principle (EU) but they make it hard to find out where to get a subscription and what that gets you. But their 18 eur per month offering is actually good with 255 eur worth of tokens included. Take care though: you need to create a specific kind of API key to avoid being charged for per token use besides your subscription.
This is on par with European stereotypes about French engineering so I’ll just go along with it :)
Is that "specific kind of API key" something that can be used in any agent, or is it Vibe CLI only?
You have to retrieve it through the Vibe interface (I used the CLI) and then you can use it with any harness.
Thanks, signed up and found it.
https://mistral.ai/pricing/ And then pick the pro plan
<mistral pm here - thanks for feedback we need to make this clearer>
> I cannot for my life figure out how to give them money
That's because the base offering is free. So you create an account, use it for free, and when you want to pay, click on "Upgrade to Pro."
I've seen some terrible B2B corporate websites that made it impossible to figure out how to give money to a company as a normal person, but I don't think this is one of them.
Well the Mistral portal is honestly kind of confusing, considering that most people just want to enter CC details and get an API key.. But it's doable!
CC details? We don't do that here, a need a QR code for ideal/wero/whatev, CC are not infra-sovereign enough
It's as easy as going to console.mistral.ai no ?
Maybe? Its impossible to tell. I'm not going to create a user in the hope that there is a subscription service in there somewhere.
> why is it so hard to give money to Mistral?
Maybe your definition of hard is much different than everyone else's...
What digital services do you pay for that don't require sign-up, or an account?
I searched for "Mistral subscription" on Kagi and this was the first link: https://mistral.ai/pricing/
You will not create a user, but want to pay?
There is a subscription in there. I'm paying.
What? It seems very doable for people who are interested in it, are you one of those people? Because it sure seems like you're trying your hardest to not find the answer yourself.
Fully agree, but I just asked Google Gemini to list Mistral's pricing and plans. Worked.
I am comparing Mistral and Claude for a company of programmer at the moment. Problem isn't so much the model but the harness and skills. Claude offers very good tools here. Mistral is behind. So a new model isn’t so interesting for me
We actually released a new harness recently - that works great. You can read more about it here https://github.com/mistralai/mistral-vibe/issues/1301
Also if I can help in any way you can reach out - isidor@mistral.ai
Unlike with Claude, you can use any harness you want with Mistral's models, including Claude Code.
> 1-trillion-parameter
I'm going to pass, but nice to know it's open now.
How much would it cost to self host this model?
Enough that if you have to ask, you can’t afford it. ;)
Open weight is always good for developers
It's nice they recently added acclaimed open models at a fair price like DeepSeek and GLM Flash, that way it's less of a gamble to try this new model out.
Isn't this inviting abliteration?
Let's hope so. Another good model for security work would be great.
Who cares? It's a crappy model and not even truly open-weight. Mistral gets absolutely mogged by Deepseek and GLM. They have never been competitive and never will be. While China and US are competing for the olympic podium, Mistral is still in the kiddy pool.
At least the FLUX vision models from Europe are pretty good.
It beats z.ai's GLM 5.3, which everyone agreed was interesting a couple of weeks ago when it was released.
It's exactly where a model trained with these resources in a modern and reasonable way should be expected to be.
I also think it demonstrates that GLM 5.3's performance is probably not primarily due to distillation but simply the performance one can get out of a a model of that size and with that architecture.
... under specific circumstances
I love how demanding the Hacker News always becomes when it comes to non-american models
Also EU fines Meta $1B: Literally stealing from this great American company!
Minor US state fines Meta $25B: After looking at Meta's financials, I guess they can afford it, no problem.
hn crowd voted for trump to get lower taxes and are now surprised people want to untangle from the US
Meow! Did nobody read the license?