We're releasing the weights of Mistral Large 4
mistral.ai
mistral.ai
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.
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.
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.
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.
Haven’t you heard the saying, "If you can’t say anything nice, don’t say anything at all"?
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).
Maybe someday.
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).
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.
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.
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.
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.
Those in the EU wanting EU-only inference should be aware Mistral put a 10% tax ("upcharge") on Regional Inference[1].
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.
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.
Guardrails on open-weight models is security theater purely to shield whoever released them from a potential PR nightmare.
"Our product won't hack you 96.6% of the time"
https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVF...
Thanks for the link, appreciated!
I have no idea how you could shrink this down to fit on any type of prosumer hardware though. I doubt it's possible.
This is the stuff I've put together and it is clearly too small for this:
https://jacquesmattheij.com/local-inference-with-large-langu...
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.
"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:
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
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.
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.
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]"
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?
[0]: https://www.frandroid.com/culture-tech/intelligence-artifici... "It will run on a laptop": Mistral confirms a lighter Large 4
100m = infant
1b = toddler
100b = age seven
1T = teenager
10T = 35 year old adult
And this tracks, because more compute allows the systems to parse more data, and learn over more cyber years. The reason that you cannot find a seven year old as smart as the average teenager is because intelligence NEEDS more data to learn more.
You can do hacks and tricks at the margins for marginal gains, but more compute for smarter models may just be a law of physics for how intelligence is made.
You need to have multiple layers in an LLM to process and refine the information. The larger the token count, etc. are the larger each layer is. The activation density of these networks is low compared to the smaller parameter models.
It may be that you need the large networks in order to learn complex tasks, but to me it seems like the big companies are throwing compute (more data, more hardware, more parameters, etc.) at the problem because that is currently easier and simpler than doing things like creating new architectures.
AFAICT, all the architecture advancements (rope, mixture of experts, reasoning, etc.) have all come from outside of the big AI companies. I think those are where we'll see the better gains, some of which are likely to make these easier to run on smaller hardware.
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.
The government can just force release, even if it's on European servers.
This is also where the words of law and pragmatic reality intercede.
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.
> ... 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"
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.
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 prefer using a European provider though so this is actually cheaper than I can find elsewhere.
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...
Weights are more like an executable binary. The source is the training data etc
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.
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.
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 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.
For example there are more than 100k modified weights on hugging face.
As a result, in practice, the only example I can think of is cracks for pirated software.
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?
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.
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.
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.
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 hope that the long term future of AI is more like Linux and less like Windows.
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.
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.
But this is neither.
Is there actually a competitive proper actually open model?
Correct me if I am wrong.
Latest versions of Kimi, Qwen, GLM aren't open, you need to pay them for commercial stuff and comply with PRC laws
[0] https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
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 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)
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.
Peer production processes (like open source) require that the inputs be specialized knowledge and relatively small amounts of capital.
"outpaced by SOTA" doesn't matter at all, the models we have are plenty good enough for most real tasks.
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.
No major Chinese or Western labs satisfy these requirements. The only vendors I can think of that do are Nvidia and Allen AI.
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.
It's just that their models were never frontier-level. By major, I meant frontier or popular.
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
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.
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.
That's not the way I remember it...
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.
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.
But this all says nothing about the original question, about whether the weights can be copyrighted.
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.
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?
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."
https://intellectual-property-helpdesk.ec.europa.eu/regional...?
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).
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.
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.
Put a JEV on the output to filter out responses that might show its Mistral
Theft is the AI way
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.
And Hashicorp's software is no longer open source, after they changed the 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.
[0] https://www.forbes.com/sites/maryroeloffs/2026/08/17/ai-comp...
Surely they'll outcompete everyone on cost alone
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.
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.
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.
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.
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.
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.
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 ?
Kimi/GLM/etc basically all have this. They're not doing this out of the kindness of their own hearts.
permissionless vs permissioned licenses
many of us in open source ascribe to the permissionless ethos, but there is a spectrum or venn
It looks extremely clear on this page: https://docs.mistral.ai/models
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.
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.
In recent years many companies mislabeled source abailable code as open source as a marketing and to build a comunity before the rug pull.
Um. Yes. That is explicitly what open source means.
Inference can be a commoditized business.
When it comes to words, popular usage defines (and redefines) the meaning.
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.
Please expand on this if you can. Why can't you? Who mandated this?
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.
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.
Alibaba (Qwen) at least has plenty of people in US offices to go arrest.
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.
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.
[0] https://digitalpolicyalert.org/change/20659-physical-and-cyb...
4 bit quants can be fine, try it and see.
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.
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.
https://artificialanalysis.ai/?models=mistral-large-4%2Cdeep...
Source: trust me bro
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?
"Weights available" does not seem to be in wide use and doesn't, to me, communicate anything different than "open weights."
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.
> 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.
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
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 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.
You shouldn't. The term existed in the community before they tried to prescriptively define it. They have no authority over the term.
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!
"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.
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:
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.
https://gitlab.com/gitlab-org/gitlab/-/blob/master/ee/LICENS...
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.
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!
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.
If it only restricted resale that would be okay in my book to call open weight. But this license is stifling.
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" ?
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.
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 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.
edit: I can see that I'm alone in thinking a company should communicate their business offers publicly, sorry for the question.
There is a subscription in there. I'm paying.
Maybe your definition of hard is much different than everyone else's...
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.
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.
This is on par with European stereotypes about French engineering so I’ll just go along with it :)
<mistral pm here - thanks for feedback we need to make this clearer>
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.
And these aren't technical discussions really, more of a risk management / political requirement thing.
If China is smart, they'll make the entire pipeline available open source.
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.
Now, perhaps Kimi 3.1 is missing from the benchmarks, and Deepseek 4.1 Flash is still has impressive!
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?
So the answer is both yes and no.
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?
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.
[*] 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
A first step on a trip which will hopefully go quite far.
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.
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.
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.
What does that mean?
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.
At least the FLUX vision models from Europe are pretty good.
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.
I'm going to pass, but nice to know it's open now.
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.
Also if I can help in any way you can reach out - isidor@mistral.ai
Minor US state fines Meta $25B: After looking at Meta's financials, I guess they can afford it, no problem.