> 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 sustainable and sufficient that incentive can be.
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
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.
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.
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.
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).
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
> 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].
> 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.
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.
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.
"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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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 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
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
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.
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 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.
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.
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?
> 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 sustainable and sufficient that incentive can be.
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
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.
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.
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/
Allen AI imposes further restrictions on their model releases, I think: https://allenai.org/responsible-use
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.
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.
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?
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.
With distillation being so effective I do wonder whether we can just crowd source open alternatives.
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.
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.
> 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.
So this is "Commons Clause" effectively?
https://commonsclause.com
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.
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?
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.
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
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.
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.
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.
No, it's open-weights, as in you can inspect them. This is not Free Software.
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.
I don't usually compile my kernel from scratch, but I'm very glad that the source is open.
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.
> 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.
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.
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.
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.
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.
Yes?
The whole reverse engineering and cracking scene is built upon this premise, you know.
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.
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.
Hence the difference between open weights and open source.
But this is neither.
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.
How does it help EU companies?
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.
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.
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.
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.
> 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
> 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.
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.
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.
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.
Kind of buried down at the end:
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!
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...
Thanks
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.
"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.
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.
Nobody cares about that number, because refusals don't lead to company-ending news headlines.
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.
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...
use MoE offloading
> 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.
Executives live in a weird bubble driven largely by consulting firms. Right now those firms are spreading FUD about Chinese models.
Theyre driven bootlickers because you dont get hired if not boot shaped.
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.
"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?
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.
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.
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.
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.
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.
Isn't this inviting abliteration?
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.
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.
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?
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.
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!
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 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 great that Europe finally has a good open-weights model.
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
How much would it cost to self host this model?
Enough that if you have to ask, you can’t afford it. ;)
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.
> 1-trillion-parameter
I'm going to pass, but nice to know it's open now.
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?
> 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.
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
PUBGM FLASH
... under specific circumstances
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.
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?
Okay, why is it not good enough?
"The model is not competitive."
Source: trust me bro
I love how demanding the Hacker News always becomes when it comes to non-american models