Nah. They are not, everyone is losing money, and just staying alive from gov subsidies. A claude PRO license is 20 usd/month, and for it to be profitable it should be somewhere around 200-300/month.
I think the Pro and other subs allow them to save big on ads (Claude Code use is required and it pushes their ads) and allow easy access to free data (CC occasionally solicits feedback and session sharing, which I'm pretty sure some users oblige). Also CC does massive prompt caching tailored to work in lockstep with their platform. With all that and who knows what else, I'd say it balances over time.
If EU were to "leverage" ASML, and that would require rather unlikely approval of Netherlands, it wouldn't end very well.
Already shipped machines can not be taken back, nor can they be stopped. Sure, you can cut the support, and buildup of new fabs will be stalled. But the other side, Chinese or USA, has many more levers to pull. Raw materials, energy (LNG), was majority of consumer goods, solar panels, batteries, semiconductors. EU doesn't mine or make most of them, and isn't nowhere close to even starting, instead, industrial base is already shrinking.
Even worse, ASML may dominate EUV, but other suppliers do very well in DUV. The moment ASML becomes unreliable, all of them will get infusion of money, as they become national security issue.
What about ASML subsidiaries in USA, like Cymer and ASML Wilton, will they take the bullet for the parent? Or will they become part of new competitor, with all the knowhow.
And China has already set domestic capacities in this area as a national priority.
Maybe true for frontier LLM, but there's plenty of space in the niches. For example, I think their TTS/STT models are pretty good, speaking from personal experience.
That's true now, but given the strategic importance of AI to nation states and the fractious relationship many have with China, I wouldn't expect the eagerness to use Chinese AI to continue indefinitely.
If you're taking it at the state level, there's no difference between Chinese and US frontier models. They go with the most accessible, which is the former. Just as trade still happens overwhelmingly between China and US despite the politics.
They’re in the EU. If they are a year or so behind and things start to plateau they’ll catch up. Americans perhaps don’t realize we can also tariff their digital goods to protect our own. There is a scenario where Americans and Chinese foot the bill and EU gets out cheap, e.g. similar to the Apple approach to AI.
there was "H" at paris at the time, they raised 200M or so, but never got so much visibility. don't know at which stage they are now or if they accomplished something
If Mistral aren't going to distill other people's large models they obviously need to train their own large models. This obviously requires money for optimization, tuning and training hardware.
Honestly being only 1 year behind makes me an optimist. You’re telling me Europe can be slightly behind with 1000x less capex and way more sustainable economics? Awesome. The world moves slower than AI progresses, I can see a scenario where 1 year isn’t a problem.
Chinese models are open, available to distill, and they also publish papers about their research. Being one year behind is a skill issue.
I think the most of the money would go to purchase hardware, But I hope they can start making adequate compensation for AI engineers and researchers to move forward.
My feeling is that its a difference in how funding works in different places. The USA will go all in with the populations pensions on a gamble, the Chinese subsidize. This way of operating is typical for the EU.
> Being one year behind is a skill issue
E.g. if you are an AI researcher in Europe you can just go to USA and make generational wealth. This is not a criticism of the EU model, but rather insane American capex effectively monopolizing.
> I think the most of the money would go to purchase hardware,
> > I think the most of the money would go to purchase hardware,
> So it's not a skill issue?
I meant by offering sovereign cloud/inference, not for training. But even if it was for training, Chinese labs have limited supply of GPUs, look what they've done. So it is a skill issue.
Also to clarify, I didn't mean European engineers' skills, I meant "you get what you pay for" as a company, that's why I hope they start offering better compensation to retain talent.
Handing in a photo-copy of a photo-copy of a photo-copy of a photo-copy of a photo-copy of a photo-copy of someone else's homework might work once or twice, but it is no way to run a (non grift) business.
Its not a grift if it works. If a Chinese corp copies at 90% quality at 50% price within a year it means your American business model was never a good one for competing in international markets. The real grift arguably is pretending that it was.
I’ve heard many folks get rejected and leave with a bad taste in their mouth. Interviews should make you feel like you’ve failed due to you not being quite there while still recommending friends to apply (“I didn’t make it but you should try to apply!” Vs “i felt like I had the privilege of even talking to the dude and then got ghosted”).
Mistral is not that bad as the comments here suggest. I am not using it as a frontier model but with simple RAG tasks and its doing great. Also OCR is pretty decent. It's a positive development that Europe is at least trying. Alternative would be: do nothing.
Which is interesting since who are the investors and what exactly are their roles? Samsung - European? BlackRock - European? Salesforce Ventures - European? Etc.
So yes it might well be
> [...] the largest equity fundraising round ever completed by a European technology company, three years after the company's launch.
The problem is that they're in a weird position between US models and Chinese models. Not as performant as US models, not as cheap as Chinese models.
Especially as Chinese models are getting better Mistral is getting less and less relevant.
It pains me because I want them to succeed, but despite them denying it I believe they'll end up restrict their activity to (1) selling hosting for Chinese models (they're already hosting GLM) and (2) selling AI-related consultant service (they're also doing that already).
I think Europe was at risk of being left without even a foot in the door. Mistral os one of few such feet. Just having some compute and know how for running Chinese models is better than nothing. Low bar I know but still. Who knows what R&D they’re doing while keeping the business running.
I’m not sure really what the winning strategy is in this weird arms race.
Mistral is odd. They have made mostly flops, boring models (Ministral 3, Mistral Small 4, Small 3, Small 3.1) together with a classic masterpiece Mistral Nemo and very good Mistral Large 2407, Mistral Small 22b, Mistral Small 3.2.
China is actively tripping the US up by pushing the open-weight strategy, which is kind of smart. They realized that they obviously will not be able to make Western companies trust Chinese companies enough to just use proprietary models by sending all their data to services hosted by Chinese companies. The US is still able to attract this trust, although it's eroding quickly in spite of recent political events and the current trust is more or less a function of old habits that need some time to change. But China correctly found that they wouldn't ever be able to compete in this way, even if they had proprietary foundation models superior to their US counterparts, so they decided to throw sticks into the spokes of the US frontier labs by releasing top open-weight models worth billions of dollars in training cost and thereby devaluing the huge proprietary investments of the US labs.
Their ‘open weight strategy’ was invented, to Chinas surprise, by the Western press after the DeepSeek event. Of course all the actual systems like Baidu, Bytedance are closed weight and basically all anyone uses. Meanwhile the advanced system are crossing security frontiers: soon only inferior models will be ‘open’ (as with OpenAI, even) the rest closed. — Or where open weighted, they will be impossible to run without a private data center, come with absurd ‘security scrutiny’ licenses like GLM is starting — or licensing requiring a cut, as with Kimi, which is basically a scheme to get western companies to do buildout for them
Europe is not China, I repeat we are not China, they build stuff. A guy from Austria ended up on Chinese national TV due to the bizarre way he had to obtain an AC unit this summer. Let's stop the larp.
I've made a personal doc tool based on Mistral OCR API at first, and then switched to Gemini Flash for the same price inline, or half price using batch API, and the difference is night and day. Mistral doesn't even come close on anything non-trivial. Longer story here: https://max.engineer/ringbinder
Mistral is an interesting AI company because they clearly have a contrarian business strategy to the other AI labs. They're also landing big customers in Europe for the right reasons. People dump on them because they're not benchmaxxxing which is pretty shortsighted - do you really want to be in a benchmark arms race with China, or do you want to make money and deploy sovereign AI compute in Europe?
I root for Mistral and hope they'll be successful, perhaps I'll buy a subscription too once they're good enough for coding aid (perhaps they are now, didn't do any test with their models recently).
First of all, they release the models' weights, perdonally I don't consider any other option as viable (no OpenAI and definitely no Anthropic, thank you).
I especially like their Vibe Chat web offer, the allowed monthly usage with a free account is incredibly generous (still have to hit a limit) and the deep research feature (5/month for free) is also valuable.
I don't know anything about the alleged regulation maxxing problems, I don't perceive them as a problem for my causal/personal usage anyway.
The gap has only been increasing, though. Devstral 2 was obviously not great compared to Claude/GPT but kind of acceptable if you were willing to compromise. There has been no real progress since then and frontier labs are massively ahead.
I don't care about benchmarks. Benchmarks show that Opus 5 is a stronger model than Fable 5 which is obviously not the case.
But I do care about capability and so far only Anthropic and, very recently with Astra, OpenAI can deliver on coding quality. And capability matters immensely. There is a world of difference between being able to do something and not being able.
A capability isn't binary. There is a massive difference between can produce an impressive demo and can reliably complete the task without a human babysitting it.
Yup, new SOTA models especially with high/xhigh/max reasoning too often overengineer solutions, good for benchmarks that usually measure task completion, bad for normal development where you don't want 'rewrite in rust and 1k LOC unit tests style solutions' when agent does mundane bug fixes.
When it comes to mundane bug fixes the value is in actually finding the cause of the bug, and I find SOTA models way outperform smaller ones here. I don't care about their output - I can write the correct 5 line patch myself once I understand what's wrong.
People fawn over AI brands now like cars and it's silly. OpenAI and Anthropic have been flipping spots for best LLM coder for the last two years and to say one is better feels silly; I've been using them both and they're very similar with different personalities. Recently Grok has become competitive in many aspects, and while I don't have much experience with Gemini it seems to come and go in terms of coding quality.
Saying only anthropic models are competitive frontier coding models is out of touch with the space imo
Genuinely interested, which ones do you think have relevance?
If I read forums and talk to people IRL most have differing opinions what model is best. Yes, for me it's pretty clear Opus is better than earlier models, but it's at least not obvious to me that the later are significant improvements.
It can be subjective, at this stage of product availability.
Personally, I hate to be frustrated by gross intellectual faults, so I did some research in the past about the best benchmarks to assess pure (simulated, apparent) intelligence. (The quality of the found benchmarks may not reflect what the models seem to do in practice, so one's experience should be compared to the raw numbers out of the benchmarks.) Good ideas emerge in the field: it was proposed and discussed on these very pages that the LLM should be able to solve "murder mysteries", for example (alongside the pattern recognition problems in which IQ tests consist, etc.).
Moreover, the LLM shall not delirate. It is an intrinsic issue with the current architectures (they do not mirror the "Foundational theory of Knowledge", which requires confidence values and relations of foundation between notions), but it is a problem with more or less presence per model. Artificial Analysis has introduced a metric for that.
Moreover again, I want an output style that works well for the purpose - must not be a clashing style like "youngspeak" ("like, awsome") or "paternalistspeak" ("when a planet likes another very much they are attracted...") or "sycophantspeak" ("your question is so deep and interesting") or "wetspeak" ("you can do it, feel this not that")... So, for example, I very much preferred Kimi k2 to gpt-oss-120b. I doubt there are benchmarks for this - "seriousspeak", "maturespeak" - but there should be.
When Apple is not racing for the frontier it's a strategy, but when it's Mistral it's a mistake.
The two companies have read the market the same.
It's always very dangerous for first movers and their investors, and the commodification of intelligence seems even more likely each time a chinese open model release. It's less exciting to do business that way, but if you're building to stand the test of time, it's wiser that way.
> When Apple is not racing for the frontier it's a strategy, but when it's Mistral it's a mistake.
I think what Mistral is doing is smart within their financial constraints, but this comparison is misleading. Mistral is an LLM company; Apple is a consumer hardware and services company.
It's smart for Apple not to join the LLM arms race, because they can just pick the cheapest supplier and let other companies take the financial losses. Mistral is in a very different situation; they are the supplier.
Mistral is a Sovereign LLM Company, it's their main product and has been from nearly the start.
They don't actually need to offer the best models, they need credibility on the tech front and the security/strategic front, and institutional clients will keep coming.
What’s the point of having “sovereign” weights that are worse than publicly available ones? Wouldn’t Europe be better off just keeping up-to-date on the Chinese releases? In the event of some schism requiring sovereign capability, or even if the Chinese pulled ahead and stopped releasing the weights, why would Europe be better off because of Mistral? (Or any country’s inferior sovereign effort make them better off?)
I think I understand the incentives that cause this to exist (it would be politically worse to say we’re just going to use Chinese models) but they are misguided. If sovereigns want to have valuable models, they should insist on world class, relevant ones like the Chinese have. Instead they embrace mediocrity in the name of sovereignty.
Not sure if you are European, but in EU it's a bit taboo to even talk about this in this manner. We like to spend a lot of money to make sure we finish last.
We know it's possible to put backdoors into LLMs, we don't have reliable ways to detect them without direct support from whoever inserted it.
Europe is less-worse-off with open weights than with… I guess it's weights-as-a-service? WaaS? The thing Anthropic and OpenAI do.
But that's not enough. As recently demonstrated, being just a few months behind with the power differential between defending with an open weight model while being attacked by a leading model, means losing absolutely.
I do not know if this holds going forward or not. It's not inconceivable that we're just about to get models that make unhackable code, using all the things software developers keep saying you need to do if you really care about security.
But anyone concerned about sovereignty can't bet the farm on this possibility. For the moment, it looks like it's a national security matter to ensure at least core state functionality (including core private sector logistics) gets the absolute best, and that the absolute best isn't going to get cut off by arbitrary whim like Mythos was.
This is true even if Europe was only defending from Russian cyberwarfare and didn't need to plan for the president of the country in which Mythos was developed, attempting to annex two NATO states.
2 things here, one is related to benchmaxxing, another to being good enough
1. Mistral isn't benchmaxxing. That doesn't mean they're better, but it does mean the benchmark gap not a good reflection of the actual gap
2. I think the "world class or nothing" framing mixes general capability with system capability.
Most deployments don't need AGI
In RL you need a model that's reliably good at one or two things, thats it.
Example:
Case of a hospital flooded in emails.
You make a system that decides which patient emails needs a human and drafts replies for the rest.
If a sovereign model is good enough at that, and you can run it on a hospital's own servers under EU jurisdiction, the frontier gap part has zero importance
Who cares about "beats DeepSeek / GPT11 / Claude Fairytale 8.9"
In the views of most Europeans, that schism already happened.
Europe was perfectly happy to rely on US software and services for decades. None of the large US tech companies would be nearly as profitable if they hadn’t had a whole continent of wealthy customers, and no competition.
I don’t think Americans are realizing yet how much has changed for us the past two years.
Can I still get Google Adsense payments when Google has my European bank account and address if there are sanctions or something like that? Can I still login to my Cloudflare account and manage a domain registered with them? Will I lose access to Outlook or Gmail? Should I rely on Claude at all?
But the problem isn't if I can bypass sanctions, use VPNs, etc, or even if a company wants or can legally do that... it's the fact I'm asking these questions at all. It's not something I'd ask 15 years ago about US companies.
It's a slow shift, but it's coming. For instance, Airbus already picked a French AWS replacement (Scaleway). It won't be all, it won't be tomorrow, but "what happens if we get tariffed / they invade Greenland" is already part of everyone's disaster planning.
The reality seems to be that for decades Europe gave only lip service to decoupling from American tech infrastructure, but in the last couple of years America has gone from being seen as a strong ally to being a major risk.
It will take time to move. Frankly as an American I hope it takes a long time and we get our shit together and rebuild our alliance with Europe. But it’s possible that the damage is not reversible in the next couple of decades and Europe will accelerate their decoupling. It’s also possible we continue to slide into imperialist authoritarianism (and Europe definitely accelerates their decoupling).
This is what I mean: Americans don’t understand the immense dividends they have enjoyed from being the defacto symbol of “progress” in the 20th and 21st centuries so far. American solutions were chosen by European customers because they were reliable trading partners with an air of modernity. Homegrown was seen as the antithesis to leapfrogging into the future.
People celebrated when McDonald’s came to their country or town. Not anymore.
It doesn't happen overnight, but the adversarial behaviour of the current administration and the tariffs really changed the perspective about America on many Europeans.
The discussion about building European alternatives had never been so mainstream. If and once they emerge, I think the shift will happen. But let's see
2 years is not a long time. What I’m telling you is that sentiments have changed dramatically, and every government and company board across the continent is taking actions to position itself according to those sentiments.
You won’t see the full effect of that for at least a decade, but that doesn’t mean it’s not happening.
The training data and knowledge is the edge, you need to build that up and maintain it. And of course mine everything you can from the American and Chinese models, like they mined everything from the internet / films / music / games etc.
Any model, even an open weight one, is fundamentally an encoding of a way of viewing the world.
What kind of "alignment" are AI labs optimizing for? Ideological alignment is the full term, self-censored into something more technological-sounding.
Every model has people behind it rating what it should and shouldn't say. Every time you ask a model and trust its answer, you become ever-so-slightly ideologically indoctrinated.
I don't want my model to reflect the views of American oligarchs or Chinese cadres. I want European values of enlightenment and humanitarianism to be the default and that's why the sovereign part is important.
Is there a specific concern you have, and what kind of performance penalty is it worth to you on say coding tasks?
Conceptually, sure I understand, but in practice it currently seems like it amounts to just using a worse model without getting anything in return. And if some hypothetical alignment to European values is important, it seems like putting the necessary effort into building a model that’s actually competitive but has this alignment is the solution, rather than accepting an inferior one.
I think for coding tasks, it probably doesn't matter as much. I'm more worried about stuff like chatbots subtly pushing or normalizing a certain world view.
I think I'm not the only one that's had this revelation, recall the Llama4 announcement. [0]
> It’s well-known that all leading LLMs have had issues with bias—specifically, they historically have leaned left when it comes to debated political and social topics. This is due to the types of training data available on the internet.
> Our goal is to remove bias from our AI models
This goal of course is self-defeatingly impossible to achieve. There is no unbiased, there's always only an unbiased relative to the bias of the observer.
Phrased differently, models weren't right leaning enough for the American oligarchy class and they publicly shared their desire to change the ideology they perpetuate.
I think an important perspective to keep in mind here is Zizek, the philosopher who's dedicated his life's work to the functioning of ideology.
> I already am eating from the trashcan all the time. The name of this trashcan is ideology. The material force of ideology - makes me not see what I'm effectively eating. It's not only our reality which enslaves us. The tragedy of our predicament - when we are within ideology, is that - when we think that we escape it into our dreams - at that point we are within ideology.
-- Slavoj Zizek
My concern is that both the Americans and the Chinese will be aligning their models more and more ideologically and that they will function as the perfect propaganda machine - surface-level objective and unthreatening, but answering every question asked from a world view decided elsewhere.
For some tasks, that won't matter and there we can use whatever model is most capable. But as we outsource more and more of our thinking to AI and use it more and more to educate impressionable young people, not having our own models will mean not getting a say in how societies views are shaped and perpetuated.
Imagine for example, the question: "What caused the French revolution?" There's many answers that might be technically correct. Which ones get emphasized is where ideology lives and gets perpetuated.
China's model for decades has been to do it cheaper and then do it better for cheaper. Just accepting this makes you an economic vassal state. We've seen this play out over decades now with other industries. I'm not blaming China for this approach, but if you want to stay relevant, then you need to compete.
You don't need to view China as some scary boogeyman who's going to use AI to attack you or w/e the current conspiracy is. They just need to continue peacefully outperforming while everyone else gets fat and lazy.
I am not sure it's correct to lump apple and Mistral's strategies together. Apple's business is selling hardware/services and their stores, but Mistral's business is AI.
Apple's strategy seems to be "wait till real business shakes out" but Mistral's strategy seems to be "go after profitable niches and avoid unwinnable fights".
Well the only place they were ever able to compete with are open models, which is by definition completely unprofitable (unless they also want to compete with Vast or Openrouter as hosting for their models or something), so that sort of makes sense from the business side of things?
I think the route to profitability is super clear and simple. To be a reasonable alternative to Chinese and US models.
I think you're probably better off using the Chinese open models right now if you're concerned about vendor lock-in or capabilities disappearing because someone's economy seems a bit fragile atm. There's no guarantee China keeps releasing open weights though, so supporting a pragmatic alternative isn't a horrible idea.
I don't think any Mistral model can match an open 30B-sized Qwen from a year ago, so right now they're not really a competitive alternative to anything at all. Except Voxtral perhaps, but that's very niche.
The other part of Apple's strategy is "lets not waste money doing all that expensive research - lets just pay them for the finished result and save money".
Whatever they do will take 2 to 4 years and I think they will do something to permanently solve their memory problem and I think it will be no different than solving their processor problem (Intel) or their long-term modem problem (Qualcomm).
Let Google spend $185 billion, and Microsoft spend $140 billion thru the end of this year on AI model building and AI hardware.
Is that really part of Apple’s strategy? They have recent examples of moving a whole bunch of stuff that requires a lot of research in-house. You’ve got stuff like their silicon, cellular modems, Vision Pro, the health business, etc.
Really Apple is pretty R&D heavy when it comes to their hardware business.
I think it’s more accurate to say that Apple sees itself as a consumer solutions provider first while a lot of the companies in the AI race are heavily focused on B2B.
The frontier AI model race is about being the first to be able to sell solutions to companies that will replace their workers and make their workers more productive.
But for B2C, the value potential just isn’t there, which is why Apple isn’t chasing it.
And now you’ve got the Mac mini/Mac Studio situation where Apple is better off selling pickaxes.
Mistral is doing one more thing: building up local know-how in the European ecosystem. This is worth more than money, you can't bootstrap an industry overnight.
There are a bunch of Europeans working in the top labs abroad. We do not lack any know how, we only lack the raw amount of capital invested in doing private research, since our companies cannot thrive and compete globally due to regulations. Mistral would be much bigger if it was founded in the US.
Or it would be gone. Or it would be in the hands of psychopaths. Europe trades off the extremes for a better middle. It doesn't produce quite as many big names, but US and EU economies are roughly the same size.
According to wikipedia, EU has a 23 trillion GDP (30 trillion PPP) and 451 million people, making ~51k GDP per person (67k PPP).
US is at 32 trillion GDP and same PPP for 342 million people - 94k per person
Ah, interesting I always thought USA’s GDP far ahead, but it’s not. Compared to EU it’s 20 vs. 30 trillions/y. Adding Switzerland, UK, etc. and it’s close.
Apple is in an entirely different market than Mistral (consumer electronics vs AI lab focusing on enterprise consulting). Which obviously means their optimal strategies are different. It doesn't really matter for Apple if it's Gemini or other model running behind their AI features. If anything it saves them a lot of money and provides a lot of flexibility.
Yea people tend to give Apple the benefit of the doubt because they’re the most successful and valuable company in human history.
Mistral is not Apple, and is not emulating their strategy. Please show me Mistral’s half a $Trillion in yearly revenue coming from consumer hardware/software.
Then I’ll agree with you that they’re taking the Apple strategy.
They aren't following the same strategy. Apple does race for the frontier in their main field: beautiful, well-integrated hardware and software. Mistral does not race for the frontier in their main field: AI creation. They're grabbing bits and bobs from people that need (or feel they need) sovereign AI. It's like if you go into making phones for the military. You aren't going for the best phones; you're going for making sure you're the only company that has the right connections to keep that customer.
AI company not keeping up with AI companies is by no definition the same as combinedhardwaresoftwareservicesentertainmentlifestyletechcompany not keeping up with AI companies
They are not the same kind of companies but they benefit both in their own way of the same market reading.
Apple is fine-tunning Gemini to customize Siri for their customers. They improve what's between the model and their customers : fine tune, inference up to the product. This is exactly what Mistral is doing, with an even more diversity of usage and needs because they are business oriented, instead of customer oriented. This is also what make them economically very efficient in comparison.
Beside Apple is still doing science experiments while Mistral is capable of releasing commercial models, albeit small and specialized.
Don't get me wrong, it's absolutely obvious that Apple is incredibly powerful, now more than ever. But the way they see the future, Mistral have a leaner trajectory.
They're enormously risk averse, and realised the infrastructure (and perhaps more important external dependency cost of building out their own training data centres). Now that the utility of the technology is clearer, and OpenAI amongst others have proven it's not only Nvidia and Google that can build a stack capable of training frontier models - I think we may find that the beast waketh from slumber.
When Microsoft is killing physical in 2023 to push for Gamepass and digital only, it's labeled as "progress and infrastructure planning", when it's Sony that does it, it's labeled as "greed and anti consumerism"
Hopefully more people get attentive to how the industry & the media works, and how US Big Tech manages to kill any form of alternative
> When Microsoft is killing physical in 2023 to push for Gamepass and digital only, it's labeled as "progress and infrastructure planning", when it's Sony that does it, it's labeled as "greed and anti consumerism"
Microsoft got massive pushback when they did it, Sony even made ad making fun of it. Selective memory much ?
Well Microsoft’s PR machine is working overtime then because I wasn’t even aware of Microsoft doing this until this HN thread whereas Sony was everywhere on social media
Apple _owns_ computers in people's pockets. There are very few businesses that can match this value. What does Mistral own? A head start at best. For the record I like Mistral and hope they succeed, but you're comparing apples to oranges.
Do they actually own it though? AFAIK they have no significant moat even in the EU (legal or competitive) and given the current accessibility to non-SOTA AI (open weight models and distillation) it's just a matter of time until Mistral gets serious competition on these points.
Siri is a minor part of the Apple ecosystem. Fundamentally, it can call into a better LLM provided by a better company. Siri's only utility is that it has access to your iData and can control your iDevices.
Apple sells devices. They have benefited a lot from their devices being good (Apple Silicon).
Mistral sells LLMs. They would benefit from having good LLMs.
Specifically we have American models which were built on extremely crappy data in insane quantities. What happens when you use same models to build a corpus of extremely high quality training data. Say for Math, coding etc. Then use that to train models. Can you get the same performance from models 10% of the size? Or 1%? Or 0.01%?
From what I've been seeing we're clearly getting to a position where models are getting "good enough" for some tasks to be really cool assistants to skilled people. And they're limited more by being extremely slow and expensive to run. What happens when they're not?
I can't see the model providers winning enough to make their valuations real.
> When Apple is not racing for the frontier it's a strategy, but when it's Mistral it's a mistake.
Ah, yes, it is really baffling that Mistral, a new European AI startup, is held to a different standard compared to a unique, 50 year old, software + hardware, 4.5 TRILLION dollar market cap company. A truly mystery of our times.
With how they’re currently being used we might as well call them bendmarks.
Every newly released model is paraded as SOTA showing peak or near peak performance on cherry-picked bendmarks the model was either fine-tuned on, or tested under specific conditions optimal for that model.
Realistically they will have to deploy the Chinese models or their finetuned versions though since their models are completely out of date and not competitive. Outside of maybe government contracts it will be hard to compete against Azure/AWS who promise to run their models in EU datacenters and not store any data since actual companies normally prefer frontier models with decent performance (cost/performance is pretty decent as well if you are fine with e.g. Luna which is massively better than anything Mistral can offer).
There are two separate issues here. One of which is very simple. That issue is where to run the models. For many companies this has to be in the EU, on EU terms. Mostly, this is not really optional from a compliance point of view. It's why all the big cloud providers have data centers in places like Frankfurt, Amsterdam, etc. and why a lot of new data centers are being built in Ireland. Of course a lot of those investments are being made by US companies. But they all have legal entities in the EU because otherwise they'd have no business here. And they can't afford to miss out on that business because it's a huge market.
The second one is about which model to run and who controls and oversees quality control. OpenAI and Anthropic seem to insist that only they can do that. But of course here in the EU we see that a bit differently. The big US based hyper-scalers are neither liked nor trusted here at this point. We don't trust the Chinese model makers much either. But with open weight models, we can at least pick different models and run them on our own terms.
Also, what most companies need is not necessarily the latest fashionable model straight from the Silicon Valley cat walk but something that will work reliably and predictably for years. Factories are not going to install the latest model in their production lines every few weeks. Same with most banks, insurers, etc. I actually know people that do business with those in relation to AI development in Germany. Companies like that are very much obsessing about self hosting their models. Sending customer data off premises is a big concern for them. They are building stuff that will be used for many years. In five years, nobody will care which model was best in autumn of 2026. But a lot of software built this year that uses AI might still be running.
You have to see Mistral's investment in that context. They could make a lot of money in the EU if they do a decent enough job. Lots of conservative companies here that are going to pick something that's good enough and then they'll be using that for many years.
"something that will work reliably and predictably for years" is not really the class of product being sold, unless you're using a fine-tuned SLM to do something like classification. The vast majority of work being done with AI unfortunately benefits from being run on the biggest/best model.
I tried them via OpenRouter. I loved their OCR. I really disliked their code generation. It was about six months ago -- so it was a geological era ago in this world. However Mistral is legally favored in Europe. In fact from my point of view there no trouble with GDPR (I live and work in Europe). I'm NOT a lawyer but I'm a technician that define itself 'privacy savy'.
>They're also landing big customers in Europe for the right reasons.
We've been migrating all of our AI automations from Gemini to Mistral because of the fear of data transfer regulations. Maybe they don't apply to us (we don't really feed personal data to AI), but we can't afford to find out.
It's been quite annoying too because the Mistral documentation and dashboards are all over the place.
Fear of fines... that's not what I would call "the right reasons".
Mistral doesn't look like it's benching at all. They're just as well funded as a lot of Chinese labs doing much more interesting work R&D-wise.
Tailoring products for compliance doesn't cut it IMO, but I'm not in their shoes.
> they clearly have a contrarian business strategy to the other AI labs.
Yeah, spot on.
I'd add they are also betting on building specialized AI's targeting narrow yet very profitable segment markets, where general AI's à la AnthroOpenAI don't work very well.
So they are parasites on sovereign blah blah blah.
And yes , Chinese models will be cheaper compared to what they offer , as well as American models will be much smarter polished . This is the only market they have , lobby sovereignty among politicians.
I really do hope they can catch up with the American models, but I think setting the Chinese models as the goal would be best. They seems to be able to create great models with low cost that are probably useful for 90% of the day to day tasks. Mistral should not focus on competing with Claude Fable or OpenAI's Astra at first but have a good EU alternative to Opus or even Sonnet. The fact that it's European will be enough to be used by a lot of companies and governments in the EU that are (trying to) move away from US tech.
"Mistral raises €3B to make sovereign, open-weight AI the technology frontier". Oh yeah, that sounds reasonable, indeed, one receives this much money just for sovereignty's sake.
Mistral just needs to good enough category think all those flash models or Qwen3.8 27b which they sadly aren't at the moment, that plus being European lab will mean that they will have very nice business. Even now these SOTA models feel too overkill for most tasks.
The question is not, will mistral be able to create the best models (seems bloody unlikely). The question is, will mistral have enough expertise to dominate its business use-case, which is the model infra/deployment market within the EU (which is more plausible, with even a 3rd, 4th, 5th best model development team).
The big US labs are naively hoping that compute power will always be their moat. I strongly suspect they are dead wrong, and newer architectures will require vastly less training data and outperform the brute force approach of US labs.
Last I saw job offers for mistral (engineering, Paris) it advertised 90k euros base salary. Not sure how they’re intending to compete with the US when even a top AI lab can’t afford to be competitive :-/
You dont even live in the Alps with 90K lmao. And if you expect top performance, you should expect top salary. Or at least something competitive. 90K for a company like Mistral is a bit embarrassing.
Yep, at least in my case (Spain), given the price of the houses and that banks are giving mortgages for 70-80% of the value, tops, you better have 100K in the bank for the down payment if you want to live in a relatively big city. So even with a good salary, is difficult to buy a house.
Beats 90% of Canadian software engineer salaries. If I wasn't tied down with 2 border collies, a chicken, a cat, and, uh, wife and kids, I'd jump to that in an instant. German quality of life is still pretty good despite all their complaining.
> Last I saw job offers for mistral (engineering, Paris) it advertised 90k euros base salary. Not sure how they’re intending to compete with the US when even a top AI lab can’t afford to be competitive :-/
Does the US position come with 5 weeks paid vacation? When the US position gets hit with layoffs, do you still have comprehensive healthcare and unemployment coverage?
Yes, EU software Eng salaries are a lot lower than their US equivalents, but quite a lot of folks are happy having a top 5% salary in their own county, without all the stresses and risks of working in the US.
As a concrete example, a few years back it leaked that Dan Abramov's (very decent) London salary was half what his US peers earned. Didn't seem to bother the man much at all...
> Does the US position come with 5 weeks paid vacation? When the US position gets hit with layoffs, do you still have comprehensive healthcare and unemployment coverage?
Work a US job for 4 months, quit, take the next 8 months for holiday/vacation, and you still save more money than a European SWE.
Hey, the 2010s are calling, they want the trope “BuT eUrOpE hAs pAiD lEaVe” back.
Europe in 2026 (France, Germany, UK, Estonia, etc). What a joke. Nobody in the world respects them.
At least the US has jobs and a military. Not as good as before, but better than fucking Europe.
Ah yeah, let me know about your healthcare too, specially after you quit that job. And unemployment. And public services. And scientific entities. And general cost of living. Good you have the military to spend a trillion a year in contractors.
> Work a US job for 4 months, quit, take the next 8 months for holiday/vacation, and you still save more money than a European SWE.
Let's assume for the sake of argument that you have a decent software job in the Bay Area, so you earn somewhere in the region of $350k. That little 4 month stint earns you just $116k (since quitting after 4 months means you'll have to return any signing bonus).
Now in the Bay Area your median mortgage is around $3,500/month, so with modest living expenses we're talking a burn of about $5,500/month, meaning you need to spend about $66,000 to cover the year. Plus $17,000 in state and federal taxes, you're now down to a cushion of 33k/year.
That's assuming, of course, that you are happy sending your kids to public school in a dodgy district (private school fees alone would wipe you out, as would the mortgages in a good public school district).
> When the US position gets hit with layoffs, do you still have comprehensive healthcare and unemployment coverage?
What does this have to do if the advertised base salary? You would pay for your healthcare and social security with huge taxes from that already meager salary, it is not like those 90k is all-taxes-paid-and-batteries-included offer.
Pay isn't the main problem with Paris, it's quality of life.
Housing near work is mostly old Haussmannian buildings that are rent capped and where demand far exceeds supply, so most of the stock is unmaintained. You either accept bad housing in the city or live in the suburbs and commute — not ideal.
To make matters worse, French companies (especially old-school ones) have a culture of presenteeism for white collar jobs, it's uncommon for people to leave before 6PM and staying late is rewarded as high engagement.
Lastly, you might consider buying and renovating a house so you can escape this dilemma, but it isn't cheap: 2-bedroom (T3) around 700k euros, that represents 20 years of frugal savings on 90k gross.
Vacation, cheap and good healthcare, and unemployment benefits are great in France, but the current government has been eroding these social benefits, and those are not unique to France for well-paid tech workers anyway.
Well said, these talking points pushed by EU politicians for decades need to stop. US white collar workers have access to pretty much the same benefits than EU ones, but have a much higher after tax compensation, even taking health care into account.
And nowadays, with the erosion of the social safety net pushed by the French government, the gap is even more narrow. There definitely are huge inequalities there, US are far from being the heartless place EU politicians like to depict.
not the good jobs. not the jobs that were comparing to mistral level equivalents in the US. i haven’t had metered PTO since 2015 and everyone i know is the same
This isn't a rebuttal of any of your point but to give a precision, the City of Paris is absurdly small, most of the Paris Area is "suburbs". The US equivalent would be if NYC was only Manhattan, and places like Brooklyn were suburbs.
Come on. We're talking about a salary that is ONE THIRD of what you'd get at AI labs in US (plus the stock and the bonus). And significantly less taxes. Are you sure that 5 weeks of vacation, what you call "comprehensive" healthcare (spoiler alert: it's not), and unemployment coverage are worth 180K/year reduction in salary (pre-taxes)?
To not talk about the toxic environment that European companies generate in general. Ie: "you should be grateful you have this job".
Look, the year is 2026, we all understand by now that unlimited PTO is a neat trick to incentivise your employees to take as little time off as possible.
> and yes you have options if laid off
Not great options. Taking California as a concrete example, you will qualify for a maximum of $450/week in unemployment payments, and if you need healthcare coverage, you will need to spend about $1,500/month on COBRA coverage for a family of 4. This is in a region where the median mortgage payment is $3,500/month.
> Does the US position come with 5 weeks paid vacation? When the US position gets hit with layoffs, do you still have comprehensive healthcare and unemployment coverage?
Maybe this appeal to mediocrity in a competitive market explains Mistral's mediocre performance.
I'm Australian. We have 4 weeks paid leave a year, comprehensive healthcare and unemployment coverage.
Most of my career I avoided taking leave, and I'd usually just get it paid out when I left. Who wants to take leave when you are doing the most interesting, most important thing you can imagine doing?
That aside, if you want to make it rational, you can think of leaves in the context of explore-exploit dilema. Probably a wider exploration would allow you to find better things to exploit. Its like the 20% innovation time that some companies offer
Most companies start at 3 weeks + 12 bank-ish holidays (or you get unlimited - which you typically have zero issues taking 5+ weeks if you're competent and your boss isn't a dickbag). Most places also give you more PTO with tenure - Amazon does, Meta doesn't.
I was going to reply with "shit americans say" until I realize you're based in Paris.
Isn't 90k a very good salary even for an expensive city like Paris? If it's not, MAN i'm out of touch with reality. In Italy 90K is great even in Milan.
Europe absolutely needs a home-grown AI lab, especially with Pax Americana looking increasingly shaky.
LLMs embody value systems, and American and European values are not the same (yes, there are overlaps, but also key differences).
More nefariously, I can also imagine LLMs that silently degrade their reasoning when used in a national security context of a non-US country.
So, Mistral may not be competitive with OpenAI and Anthropic, but in many contexts that doesn’t matter. And, perhaps this gap could be closed with more funding (the three billion funding figure is a rounding error next to US labs). I’m sort of surprised that the EU isn’t stepping in to support them.
> LLMs embody value systems, and American and European values are not the same (yes, there are overlaps, but also key differences).
There are also some areas where values across Europe are quite divergent. Think LGBT rights and social acceptance, religion & secularism, immigration & multiculturalism.
And countries don’t fall into neat “liberal west vs conservative east”. Spain is exceptionally liberal on LGBT issues while remaining more religious than some northern countries; Denmark is socially liberal but has adopted relatively restrictive immigration policies.
> Spain is exceptionally liberal on LGBT issues while remaining more religious than some northern countries
When you move here, "more religious" turns out to pretty much be window dressing. Yes, they still celebrate a bunch of the old catholic holidays, parading saints around on feast days, but apart from a handful of older folks, nobody actually believes.
Particularly when compared to the US, where a large minority is still going around loudly thumping bibles, Spain is a very secular country.
>I’m sort of surprised that the EU isn’t stepping in to support them.
On a similar note: Why does it have to be the EU to step up?
Why doesn't EU rather speed up making VC investments more attractive, so that EU and banks don't do the majority of investing?
*I don't have answers to these questions. It just frustrates me how many investments here come from politicians and banks, rather than from investors, people, and companies.
It’s a different culture with different rules, if it was that easy it would have been done already. See my other comment about lack of budget at a federal level.
It's possible to be a waste of European money that could be better used elsewhere, though. I think the same about LeCun's company sucking up the little funding here.
Imagine the fund were distributing to several teams, but let's say Z.ai and Deepseek were european and among them. What percentage would be allocated to Mistral in that case? OK, the other two aren't, but in the same vein, should the EU fund be saving its powder until such a project shows up?
I don't know. Government funds have a unique way of being often wrong, so this may be the kiss of death. But it is a sign that the EU is helping Mistral.
> LLMs embody value systems, and American and European values are not the same
Your values and American values might not be the same, but to say even most of Europe feels the same way is a big stretch. And not even the US has very many shared values anymore.
If we’re being honest, Europe doesn’t actually have any common value system outside of whatever is momentarily trendy in the urban monoculture, which is why it refuses to work together on most things and is currently being torn apart at the seams (see the rise of nationalist far right parties in most states).
The EU are a collection of states where the average citizen can’t even communicate with their neighbor in a common language beyond the level of a 6 year old.
Can you offer a believable rebuttal to any of my statements?
AfD just had a historic victory in Germany, France is leaning the same direction, etc. Just those two are 40% of the EU economy, and do you think them both turning ethno-nationalist and right wing means the EU project is going to get stronger or weaker? More shared values, or less?
The truth is the EU didn't go far enough, and created a feckless bike-shedding bureaucracy of papercuts. If the EU were to truly unite around shared values, we could achieve something. But the fundamental issue is the EU doesn't have these shared values. The US had the benefit of starting from 0 and attracting only likeminded people to grow.
In Europe every country still identifies with their early 1900s national romantic movement definition of who they are (which was itself a construct). The lack of shared values is the problem. We can't even get most of the EU on a common language.
But anytime you bring his name up some right wing conspiracy nutjob will call you an evil globalist. And the far left nutjobs (more common here) will call you an evil capitalist/fascist for trying to compete in global markets.
Europe has no budget of their own like the USA and China do, because there is no debt nor taxation at the European level. That’s the reason for the complication you are noticing.
What do you mean by "Europe has no budget of their own like the USA and China do"? The EU was specifically mentioned and the EU does have a budget, debt, and income:
> Europe absolutely needs a home-grown AI lab, especially with Pax Americana looking increasingly shaky.
The only problem (at least in the LLM space) is that you can do more in Europe by just getting the best Chinese Open weights model (do a finetune if you really want) and do more for cheaper than using Mistral.
10-20x more still needed, and then somewhere to build a couple DCs. Fingers crossed they make it, neither US nor Chinese labs can be trusted, even with open weights.
Mistral has solid OCR, STT and TTS models and I would love to support them by switching with all of our business workloads to Mistral... but their LLM models are sadly not competitive at all. In our business benchmarks their Mistral Medium 3.5 with reasoning is worse than Gemma 4 31B and Glimmer 30B. It's a 128B dense model that's priced accordingly! Mistral Small 4 is way worse than Gemma 4 26B A4B. I applaude the effort that they release those models as open weight but Gemma 4 models are currently way easier to run with more tok/s and less hardware. Their API pricing is just insane for what you get. But I guess enterprise customers don't care about it, this is why they are probably not lowering it.
> Mistral Small 4 is way worse than Gemma 4 26B A4B
Depends for what purpose? I found large Mistrals are massively better than Gemma 4 at creative writing: have more natural tone, better consistency than 26B as it is MoE.
Mistral Small 4 is also a MoE model, with way more params, so I would expect it to perform better. Our benchmark involves around 10% communication and writing in German and English and Gemma 4 26B A4B beats it although creative writing is the only category in our benchmark which might be a bit subjective.
It makes no sense to train frontier models from scratch anymore. The best frontier models are only a half year ahead of Chinese open models. In this regard Anthropic and OpenAI are also in a bad spot when they waste so much compute on training models.
An important factor is that fine tuning existing open models is incredible cheap. You can easily change any cultural biases if you want a model to be 'sovereign'. And Mistral could combine that with their custom data sets for their enterprise customer needs. Mistral still trains their own models, but they also seem to offer fine tuning existing models.
With model weights being commoditized, another differentiator could be deploying efficient inference chips, especially if you combine it with a developer ecosystem for vendor lock-in. That is why it is interesting that both Samsung and ASML are investors, since they are companies that could make a difference in this area.
It only makes sense to train a frontier model if you are trying a different architecture to one that is available from an existing frontier model. This is because the different model architecture will learn the weights differently.
It may make sense to train a frontier model on an existing architecture if the base model is not available and the instruction trained version doesn't fit with what you want. There are techniques like ablation, but those could have other effects on the model, and there can still be lingering effects of the instruction training in the model that surface less frequently (e.g. on an input not covered by the ablation training).
Otherwise, fine tuning is definitely the way to go. However, you need to be careful not to over-tune the model such that it is only tuned to the data you are training it on.
On a higher level it might make sense to build the expertise that comes with base training.
I dont know enough about the process to estimate these gains, but china has been doing it in manufacturing for decades.
All the money in the world is useless when no one knows how to do the thing
324 comments
[ 7.6 ms ] story [ 108 ms ] threadI don't trust anything the antichrist invests in
The bubble is going to burst soon.
In the EU it seems we are very much at risk of being cut off from frontier AI if the US government should decide to do so.
Already shipped machines can not be taken back, nor can they be stopped. Sure, you can cut the support, and buildup of new fabs will be stalled. But the other side, Chinese or USA, has many more levers to pull. Raw materials, energy (LNG), was majority of consumer goods, solar panels, batteries, semiconductors. EU doesn't mine or make most of them, and isn't nowhere close to even starting, instead, industrial base is already shrinking.
Even worse, ASML may dominate EUV, but other suppliers do very well in DUV. The moment ASML becomes unreliable, all of them will get infusion of money, as they become national security issue.
What about ASML subsidiaries in USA, like Cymer and ASML Wilton, will they take the bullet for the parent? Or will they become part of new competitor, with all the knowhow.
And China has already set domestic capacities in this area as a national priority.
From what I've seen in the past few months, Mistral is the unique European lab still trying to compete (I wouldn't count Poolside as EU).
I think the most of the money would go to purchase hardware, But I hope they can start making adequate compensation for AI engineers and researchers to move forward.
> Being one year behind is a skill issue
E.g. if you are an AI researcher in Europe you can just go to USA and make generational wealth. This is not a criticism of the EU model, but rather insane American capex effectively monopolizing.
> I think the most of the money would go to purchase hardware,
So it's not a skill issue?
> So it's not a skill issue?
I meant by offering sovereign cloud/inference, not for training. But even if it was for training, Chinese labs have limited supply of GPUs, look what they've done. So it is a skill issue.
Also to clarify, I didn't mean European engineers' skills, I meant "you get what you pay for" as a company, that's why I hope they start offering better compensation to retain talent.
Handing in a photo-copy of a photo-copy of a photo-copy of a photo-copy of a photo-copy of a photo-copy of someone else's homework might work once or twice, but it is no way to run a (non grift) business.
Not even a phone call :(
Anyway, the interview process was so janky it decreased my confidence in their success.
Not easy as it sounds yes, but it is better than throwing your CV into the ATS void 1000 times (the wrong way to apply for a job btw)
Mistral have no moat as it seems except some 'regulatory' moat.
Better than throwing CVs everywhere with no response for years.
From both the hiring and enterprise adoption sides, it’s easy (and disappointing) to see why the US/China seems so far ahead.
Which is interesting since who are the investors and what exactly are their roles? Samsung - European? BlackRock - European? Salesforce Ventures - European? Etc.
So yes it might well be
> [...] the largest equity fundraising round ever completed by a European technology company, three years after the company's launch.
but the money isn't European.
Especially as Chinese models are getting better Mistral is getting less and less relevant.
It pains me because I want them to succeed, but despite them denying it I believe they'll end up restrict their activity to (1) selling hosting for Chinese models (they're already hosting GLM) and (2) selling AI-related consultant service (they're also doing that already).
So they’ll probably make more money creating PowerPoints with ChatGPT than they will trying to compete with the US and China.
Which would be the most European outcome ever.
I’m not sure really what the winning strategy is in this weird arms race.
Sovereign, data privacy, these are their strengths. Or even being not US and not Chinese; this is in itself an advantage nowadays. (Which is crazy)
That is absolutely not "do nothing".
First of all, they release the models' weights, perdonally I don't consider any other option as viable (no OpenAI and definitely no Anthropic, thank you).
I especially like their Vibe Chat web offer, the allowed monthly usage with a free account is incredibly generous (still have to hit a limit) and the deep research feature (5/month for free) is also valuable.
I don't know anything about the alleged regulation maxxing problems, I don't perceive them as a problem for my causal/personal usage anyway.
The gap has only been increasing, though. Devstral 2 was obviously not great compared to Claude/GPT but kind of acceptable if you were willing to compromise. There has been no real progress since then and frontier labs are massively ahead.
But I do care about capability and so far only Anthropic and, very recently with Astra, OpenAI can deliver on coding quality. And capability matters immensely. There is a world of difference between being able to do something and not being able.
Saying only anthropic models are competitive frontier coding models is out of touch with the space imo
You must care about good benchmarks (identify those that have relevance).
If I read forums and talk to people IRL most have differing opinions what model is best. Yes, for me it's pretty clear Opus is better than earlier models, but it's at least not obvious to me that the later are significant improvements.
It can be subjective, at this stage of product availability.
Personally, I hate to be frustrated by gross intellectual faults, so I did some research in the past about the best benchmarks to assess pure (simulated, apparent) intelligence. (The quality of the found benchmarks may not reflect what the models seem to do in practice, so one's experience should be compared to the raw numbers out of the benchmarks.) Good ideas emerge in the field: it was proposed and discussed on these very pages that the LLM should be able to solve "murder mysteries", for example (alongside the pattern recognition problems in which IQ tests consist, etc.).
Moreover, the LLM shall not delirate. It is an intrinsic issue with the current architectures (they do not mirror the "Foundational theory of Knowledge", which requires confidence values and relations of foundation between notions), but it is a problem with more or less presence per model. Artificial Analysis has introduced a metric for that.
Moreover again, I want an output style that works well for the purpose - must not be a clashing style like "youngspeak" ("like, awsome") or "paternalistspeak" ("when a planet likes another very much they are attracted...") or "sycophantspeak" ("your question is so deep and interesting") or "wetspeak" ("you can do it, feel this not that")... So, for example, I very much preferred Kimi k2 to gpt-oss-120b. I doubt there are benchmarks for this - "seriousspeak", "maturespeak" - but there should be.
You crazy to think that Fable and Astra capabilities is fake
The two companies have read the market the same.
It's always very dangerous for first movers and their investors, and the commodification of intelligence seems even more likely each time a chinese open model release. It's less exciting to do business that way, but if you're building to stand the test of time, it's wiser that way.
I think what Mistral is doing is smart within their financial constraints, but this comparison is misleading. Mistral is an LLM company; Apple is a consumer hardware and services company.
It's smart for Apple not to join the LLM arms race, because they can just pick the cheapest supplier and let other companies take the financial losses. Mistral is in a very different situation; they are the supplier.
They don't actually need to offer the best models, they need credibility on the tech front and the security/strategic front, and institutional clients will keep coming.
I think I understand the incentives that cause this to exist (it would be politically worse to say we’re just going to use Chinese models) but they are misguided. If sovereigns want to have valuable models, they should insist on world class, relevant ones like the Chinese have. Instead they embrace mediocrity in the name of sovereignty.
Europe is less-worse-off with open weights than with… I guess it's weights-as-a-service? WaaS? The thing Anthropic and OpenAI do.
But that's not enough. As recently demonstrated, being just a few months behind with the power differential between defending with an open weight model while being attacked by a leading model, means losing absolutely.
I do not know if this holds going forward or not. It's not inconceivable that we're just about to get models that make unhackable code, using all the things software developers keep saying you need to do if you really care about security.
But anyone concerned about sovereignty can't bet the farm on this possibility. For the moment, it looks like it's a national security matter to ensure at least core state functionality (including core private sector logistics) gets the absolute best, and that the absolute best isn't going to get cut off by arbitrary whim like Mythos was.
This is true even if Europe was only defending from Russian cyberwarfare and didn't need to plan for the president of the country in which Mythos was developed, attempting to annex two NATO states.
1. Mistral isn't benchmaxxing. That doesn't mean they're better, but it does mean the benchmark gap not a good reflection of the actual gap
2. I think the "world class or nothing" framing mixes general capability with system capability. Most deployments don't need AGI In RL you need a model that's reliably good at one or two things, thats it. Example: Case of a hospital flooded in emails. You make a system that decides which patient emails needs a human and drafts replies for the rest. If a sovereign model is good enough at that, and you can run it on a hospital's own servers under EU jurisdiction, the frontier gap part has zero importance
Who cares about "beats DeepSeek / GPT11 / Claude Fairytale 8.9"
Does Mistral have material market share for any application, including anything that would fall under item 2 above?
In the views of most Europeans, that schism already happened.
Europe was perfectly happy to rely on US software and services for decades. None of the large US tech companies would be nearly as profitable if they hadn’t had a whole continent of wealthy customers, and no competition.
I don’t think Americans are realizing yet how much has changed for us the past two years.
Never had to do it before. That's how much it has changed.
You don't think Ukrainians can't buy stuff from Russia (and vice-versa)?
But the problem isn't if I can bypass sanctions, use VPNs, etc, or even if a company wants or can legally do that... it's the fact I'm asking these questions at all. It's not something I'd ask 15 years ago about US companies.
It will take time to move. Frankly as an American I hope it takes a long time and we get our shit together and rebuild our alliance with Europe. But it’s possible that the damage is not reversible in the next couple of decades and Europe will accelerate their decoupling. It’s also possible we continue to slide into imperialist authoritarianism (and Europe definitely accelerates their decoupling).
This is what I mean: Americans don’t understand the immense dividends they have enjoyed from being the defacto symbol of “progress” in the 20th and 21st centuries so far. American solutions were chosen by European customers because they were reliable trading partners with an air of modernity. Homegrown was seen as the antithesis to leapfrogging into the future.
People celebrated when McDonald’s came to their country or town. Not anymore.
The discussion about building European alternatives had never been so mainstream. If and once they emerge, I think the shift will happen. But let's see
You won’t see the full effect of that for at least a decade, but that doesn’t mean it’s not happening.
In two more, possibly less, the current administration will be gone. The Trump agenda will be dead in the water by the end of this year.
While I basically agree with you, I wonder how quickly people will forget.
Frankly this already started under GW Bush from my perspective and just has gradually gotten worse.
I don’t think you can unboil a frog that quickly. This will take many years to resolve / build up trust again.
What kind of "alignment" are AI labs optimizing for? Ideological alignment is the full term, self-censored into something more technological-sounding.
Every model has people behind it rating what it should and shouldn't say. Every time you ask a model and trust its answer, you become ever-so-slightly ideologically indoctrinated.
I don't want my model to reflect the views of American oligarchs or Chinese cadres. I want European values of enlightenment and humanitarianism to be the default and that's why the sovereign part is important.
Conceptually, sure I understand, but in practice it currently seems like it amounts to just using a worse model without getting anything in return. And if some hypothetical alignment to European values is important, it seems like putting the necessary effort into building a model that’s actually competitive but has this alignment is the solution, rather than accepting an inferior one.
I think I'm not the only one that's had this revelation, recall the Llama4 announcement. [0]
> It’s well-known that all leading LLMs have had issues with bias—specifically, they historically have leaned left when it comes to debated political and social topics. This is due to the types of training data available on the internet.
> Our goal is to remove bias from our AI models
This goal of course is self-defeatingly impossible to achieve. There is no unbiased, there's always only an unbiased relative to the bias of the observer.
Phrased differently, models weren't right leaning enough for the American oligarchy class and they publicly shared their desire to change the ideology they perpetuate.
I think an important perspective to keep in mind here is Zizek, the philosopher who's dedicated his life's work to the functioning of ideology.
> I already am eating from the trashcan all the time. The name of this trashcan is ideology. The material force of ideology - makes me not see what I'm effectively eating. It's not only our reality which enslaves us. The tragedy of our predicament - when we are within ideology, is that - when we think that we escape it into our dreams - at that point we are within ideology.
-- Slavoj Zizek
My concern is that both the Americans and the Chinese will be aligning their models more and more ideologically and that they will function as the perfect propaganda machine - surface-level objective and unthreatening, but answering every question asked from a world view decided elsewhere.
For some tasks, that won't matter and there we can use whatever model is most capable. But as we outsource more and more of our thinking to AI and use it more and more to educate impressionable young people, not having our own models will mean not getting a say in how societies views are shaped and perpetuated.
Imagine for example, the question: "What caused the French revolution?" There's many answers that might be technically correct. Which ones get emphasized is where ideology lives and gets perpetuated.
[0] https://ai.meta.com/blog/llama-4-multimodal-intelligence/
You don't need to view China as some scary boogeyman who's going to use AI to attack you or w/e the current conspiracy is. They just need to continue peacefully outperforming while everyone else gets fat and lazy.
Apple's strategy seems to be "wait till real business shakes out" but Mistral's strategy seems to be "go after profitable niches and avoid unwinnable fights".
I think you're probably better off using the Chinese open models right now if you're concerned about vendor lock-in or capabilities disappearing because someone's economy seems a bit fragile atm. There's no guarantee China keeps releasing open weights though, so supporting a pragmatic alternative isn't a horrible idea.
they could seriously do something about it instead of waiting.
Let Google spend $185 billion, and Microsoft spend $140 billion thru the end of this year on AI model building and AI hardware.
Really Apple is pretty R&D heavy when it comes to their hardware business.
I think it’s more accurate to say that Apple sees itself as a consumer solutions provider first while a lot of the companies in the AI race are heavily focused on B2B.
The frontier AI model race is about being the first to be able to sell solutions to companies that will replace their workers and make their workers more productive.
But for B2C, the value potential just isn’t there, which is why Apple isn’t chasing it.
And now you’ve got the Mac mini/Mac Studio situation where Apple is better off selling pickaxes.
So even worse
According to wikipedia, EU has a 23 trillion GDP (30 trillion PPP) and 451 million people, making ~51k GDP per person (67k PPP). US is at 32 trillion GDP and same PPP for 342 million people - 94k per person
Straight out of "The Art of War" by Sun Tzu.
Mistral is not Apple, and is not emulating their strategy. Please show me Mistral’s half a $Trillion in yearly revenue coming from consumer hardware/software.
Then I’ll agree with you that they’re taking the Apple strategy.
Apple is a $4.7 trillion company selling computers and iPhones. How many computers and iPhones is Mistral selling?
In your comparison Apple is Apple while Mistral is orange.
Apple is fine-tunning Gemini to customize Siri for their customers. They improve what's between the model and their customers : fine tune, inference up to the product. This is exactly what Mistral is doing, with an even more diversity of usage and needs because they are business oriented, instead of customer oriented. This is also what make them economically very efficient in comparison.
Beside Apple is still doing science experiments while Mistral is capable of releasing commercial models, albeit small and specialized.
Don't get me wrong, it's absolutely obvious that Apple is incredibly powerful, now more than ever. But the way they see the future, Mistral have a leaner trajectory.
In 2010s, they were among the "greats" of consumer AI. I don't think their actions now are "strategy" and not "skill issue".
It's the same with gaming
When Microsoft is killing physical in 2023 to push for Gamepass and digital only, it's labeled as "progress and infrastructure planning", when it's Sony that does it, it's labeled as "greed and anti consumerism"
Hopefully more people get attentive to how the industry & the media works, and how US Big Tech manages to kill any form of alternative
Microsoft got massive pushback when they did it, Sony even made ad making fun of it. Selective memory much ?
-2013 DRM/always online drama (reversed pre launch)
-2023 shutdown of its physical release division, which is what triggered shift to digital only titles we are seeing now
Different year, different mechanism, different reception and different press coverage
Independence from the USA's CLOUD Act, secret FISA courts, and ICC-style disabling of important services.*
*…and the Chinese equivalents.
Actually, he’s comparing apples to mistrals.
I’ll get my coat.
*Please verify results with human based weather perception
Siri is a minor part of the Apple ecosystem. Fundamentally, it can call into a better LLM provided by a better company. Siri's only utility is that it has access to your iData and can control your iDevices.
Apple sells devices. They have benefited a lot from their devices being good (Apple Silicon).
Mistral sells LLMs. They would benefit from having good LLMs.
Which is smart move
From what I've been seeing we're clearly getting to a position where models are getting "good enough" for some tasks to be really cool assistants to skilled people. And they're limited more by being extremely slow and expensive to run. What happens when they're not?
I can't see the model providers winning enough to make their valuations real.
Also: When China does it, it is protectionism; when Europe does it, it is sovereignty.
Ah, yes, it is really baffling that Mistral, a new European AI startup, is held to a different standard compared to a unique, 50 year old, software + hardware, 4.5 TRILLION dollar market cap company. A truly mystery of our times.
Every newly released model is paraded as SOTA showing peak or near peak performance on cherry-picked bendmarks the model was either fine-tuned on, or tested under specific conditions optimal for that model.
The second one is about which model to run and who controls and oversees quality control. OpenAI and Anthropic seem to insist that only they can do that. But of course here in the EU we see that a bit differently. The big US based hyper-scalers are neither liked nor trusted here at this point. We don't trust the Chinese model makers much either. But with open weight models, we can at least pick different models and run them on our own terms.
Also, what most companies need is not necessarily the latest fashionable model straight from the Silicon Valley cat walk but something that will work reliably and predictably for years. Factories are not going to install the latest model in their production lines every few weeks. Same with most banks, insurers, etc. I actually know people that do business with those in relation to AI development in Germany. Companies like that are very much obsessing about self hosting their models. Sending customer data off premises is a big concern for them. They are building stuff that will be used for many years. In five years, nobody will care which model was best in autumn of 2026. But a lot of software built this year that uses AI might still be running.
You have to see Mistral's investment in that context. They could make a lot of money in the EU if they do a decent enough job. Lots of conservative companies here that are going to pick something that's good enough and then they'll be using that for many years.
I'm not sure that's true amongst most large enterprise companies.
> They could make a lot of money in the EU if they do a decent enough job
They could but unfortunately there have only ever been a small handful of European tech companies which got anywhere close to that
which is a market Chinese labs won't get into.
they only other company they compete with is probably palantir in that regard.
We've been migrating all of our AI automations from Gemini to Mistral because of the fear of data transfer regulations. Maybe they don't apply to us (we don't really feed personal data to AI), but we can't afford to find out.
It's been quite annoying too because the Mistral documentation and dashboards are all over the place.
Fear of fines... that's not what I would call "the right reasons".
What’s contrarian? Dont they also sell API and subscription like every other lab?
Their own models seems like years old OpenAI models, hallucinations all over the place and coding was crazy slow.
I can only say two good things about them:
- their hosted glm model was fast - because I canceled within 14 days they gave me a full refund.
Yeah, spot on.
I'd add they are also betting on building specialized AI's targeting narrow yet very profitable segment markets, where general AI's à la AnthroOpenAI don't work very well.
Because this forum is sponsored by Claude and Anslopic.
Anything against the narrative is attacked.
Used mistral 7b locally for years - it’s fine.
Their web version is like a slightly worse Gemini - also fine
And yes , Chinese models will be cheaper compared to what they offer , as well as American models will be much smarter polished . This is the only market they have , lobby sovereignty among politicians.
They would if they could, which means they can‘t, even though they want to.
I want my money back.
Mistral's revenue is mostly B2B, and that's much more difficult to move.
this comment is insane...
Thats quite competitive for a base salary, as high as max ICT5 Staff base at Apple Munich.
Does the US position come with 5 weeks paid vacation? When the US position gets hit with layoffs, do you still have comprehensive healthcare and unemployment coverage?
Yes, EU software Eng salaries are a lot lower than their US equivalents, but quite a lot of folks are happy having a top 5% salary in their own county, without all the stresses and risks of working in the US.
As a concrete example, a few years back it leaked that Dan Abramov's (very decent) London salary was half what his US peers earned. Didn't seem to bother the man much at all...
Work a US job for 4 months, quit, take the next 8 months for holiday/vacation, and you still save more money than a European SWE.
Hey, the 2010s are calling, they want the trope “BuT eUrOpE hAs pAiD lEaVe” back.
Europe in 2026 (France, Germany, UK, Estonia, etc). What a joke. Nobody in the world respects them.
At least the US has jobs and a military. Not as good as before, but better than fucking Europe.
Let's assume for the sake of argument that you have a decent software job in the Bay Area, so you earn somewhere in the region of $350k. That little 4 month stint earns you just $116k (since quitting after 4 months means you'll have to return any signing bonus).
Now in the Bay Area your median mortgage is around $3,500/month, so with modest living expenses we're talking a burn of about $5,500/month, meaning you need to spend about $66,000 to cover the year. Plus $17,000 in state and federal taxes, you're now down to a cushion of 33k/year.
That's assuming, of course, that you are happy sending your kids to public school in a dodgy district (private school fees alone would wipe you out, as would the mortgages in a good public school district).
The math doesn't math here chief
Yes, absolutely, the problem is these are not the people you want to hire.
And the people you do want to hire are either already in US or work for a US company remotely for 3x the salary.
What does this have to do if the advertised base salary? You would pay for your healthcare and social security with huge taxes from that already meager salary, it is not like those 90k is all-taxes-paid-and-batteries-included offer.
Housing near work is mostly old Haussmannian buildings that are rent capped and where demand far exceeds supply, so most of the stock is unmaintained. You either accept bad housing in the city or live in the suburbs and commute — not ideal.
To make matters worse, French companies (especially old-school ones) have a culture of presenteeism for white collar jobs, it's uncommon for people to leave before 6PM and staying late is rewarded as high engagement.
Lastly, you might consider buying and renovating a house so you can escape this dilemma, but it isn't cheap: 2-bedroom (T3) around 700k euros, that represents 20 years of frugal savings on 90k gross.
Vacation, cheap and good healthcare, and unemployment benefits are great in France, but the current government has been eroding these social benefits, and those are not unique to France for well-paid tech workers anyway.
And nowadays, with the erosion of the social safety net pushed by the French government, the gap is even more narrow. There definitely are huge inequalities there, US are far from being the heartless place EU politicians like to depict.
1. all good tech jobs in the US have unlimited pto
2. all good tech jobs in the US have good healthcare plans that will continue after being let go until you get a new job
3. you will be able to claim unemployment, but a good tech company in the US will also give you a severance
Its usually 15 days paid time off, and unlimited unpaid.
To not talk about the toxic environment that European companies generate in general. Ie: "you should be grateful you have this job".
that buys you a lot of dentists visits and vacation funds
Look, the year is 2026, we all understand by now that unlimited PTO is a neat trick to incentivise your employees to take as little time off as possible.
> and yes you have options if laid off
Not great options. Taking California as a concrete example, you will qualify for a maximum of $450/week in unemployment payments, and if you need healthcare coverage, you will need to spend about $1,500/month on COBRA coverage for a family of 4. This is in a region where the median mortgage payment is $3,500/month.
Maybe this appeal to mediocrity in a competitive market explains Mistral's mediocre performance.
I'm Australian. We have 4 weeks paid leave a year, comprehensive healthcare and unemployment coverage.
Most of my career I avoided taking leave, and I'd usually just get it paid out when I left. Who wants to take leave when you are doing the most interesting, most important thing you can imagine doing?
That aside, if you want to make it rational, you can think of leaves in the context of explore-exploit dilema. Probably a wider exploration would allow you to find better things to exploit. Its like the 20% innovation time that some companies offer
It's relatively common to get 4+ weeks paid in GOOD tech jobs in the US - of which these would be considered.
I didn't have that much at either Amazon or Meta.
https://www.levels.fyi/benefits/PTO-Vacation-Personal-Days/#...
Most companies start at 3 weeks + 12 bank-ish holidays (or you get unlimited - which you typically have zero issues taking 5+ weeks if you're competent and your boss isn't a dickbag). Most places also give you more PTO with tenure - Amazon does, Meta doesn't.
It probably means they don't, they just do their own thing how they see fit.
LLMs embody value systems, and American and European values are not the same (yes, there are overlaps, but also key differences).
More nefariously, I can also imagine LLMs that silently degrade their reasoning when used in a national security context of a non-US country.
So, Mistral may not be competitive with OpenAI and Anthropic, but in many contexts that doesn’t matter. And, perhaps this gap could be closed with more funding (the three billion funding figure is a rounding error next to US labs). I’m sort of surprised that the EU isn’t stepping in to support them.
There are also some areas where values across Europe are quite divergent. Think LGBT rights and social acceptance, religion & secularism, immigration & multiculturalism.
And countries don’t fall into neat “liberal west vs conservative east”. Spain is exceptionally liberal on LGBT issues while remaining more religious than some northern countries; Denmark is socially liberal but has adopted relatively restrictive immigration policies.
When you move here, "more religious" turns out to pretty much be window dressing. Yes, they still celebrate a bunch of the old catholic holidays, parading saints around on feast days, but apart from a handful of older folks, nobody actually believes.
Particularly when compared to the US, where a large minority is still going around loudly thumping bibles, Spain is a very secular country.
To Caesar what is Caesar’s, to God…
On a similar note: Why does it have to be the EU to step up?
Why doesn't EU rather speed up making VC investments more attractive, so that EU and banks don't do the majority of investing?
*I don't have answers to these questions. It just frustrates me how many investments here come from politicians and banks, rather than from investors, people, and companies.
Nah, it's more that there's no Capital Markets Union, so there's just ~30 smaller pots of money scattered across each country.
that is being replaced with Pax Silica
https://www.state.gov/pax-silica
It is! The second largest investor in this round is Scaleup Europe Fund:
https://eic.ec.europa.eu/eic-fund/scaleup-europe-fund_en
Your values and American values might not be the same, but to say even most of Europe feels the same way is a big stretch. And not even the US has very many shared values anymore.
If we’re being honest, Europe doesn’t actually have any common value system outside of whatever is momentarily trendy in the urban monoculture, which is why it refuses to work together on most things and is currently being torn apart at the seams (see the rise of nationalist far right parties in most states).
The EU are a collection of states where the average citizen can’t even communicate with their neighbor in a common language beyond the level of a 6 year old.
AfD just had a historic victory in Germany, France is leaning the same direction, etc. Just those two are 40% of the EU economy, and do you think them both turning ethno-nationalist and right wing means the EU project is going to get stronger or weaker? More shared values, or less?
The truth is the EU didn't go far enough, and created a feckless bike-shedding bureaucracy of papercuts. If the EU were to truly unite around shared values, we could achieve something. But the fundamental issue is the EU doesn't have these shared values. The US had the benefit of starting from 0 and attracting only likeminded people to grow.
In Europe every country still identifies with their early 1900s national romantic movement definition of who they are (which was itself a construct). The lack of shared values is the problem. We can't even get most of the EU on a common language.
There were people like Kalgeri who had a vision for a 'united states of europe,' which arguably would have been better: https://en.wikipedia.org/wiki/Richard_von_Coudenhove-Kalergi
But anytime you bring his name up some right wing conspiracy nutjob will call you an evil globalist. And the far left nutjobs (more common here) will call you an evil capitalist/fascist for trying to compete in global markets.
I really, really wish that this was true, but basically no-one is pushing for euro bonds (even though they'd be a great idea).
https://european-union.europa.eu/institutions-law-budget/bud...
The only problem (at least in the LLM space) is that you can do more in Europe by just getting the best Chinese Open weights model (do a finetune if you really want) and do more for cheaper than using Mistral.
Maybe we need to abolish copyright?
Depends for what purpose? I found large Mistrals are massively better than Gemma 4 at creative writing: have more natural tone, better consistency than 26B as it is MoE.
Perhaps Mistral has too high moral standards to keep up?
An important factor is that fine tuning existing open models is incredible cheap. You can easily change any cultural biases if you want a model to be 'sovereign'. And Mistral could combine that with their custom data sets for their enterprise customer needs. Mistral still trains their own models, but they also seem to offer fine tuning existing models.
With model weights being commoditized, another differentiator could be deploying efficient inference chips, especially if you combine it with a developer ecosystem for vendor lock-in. That is why it is interesting that both Samsung and ASML are investors, since they are companies that could make a difference in this area.
It may make sense to train a frontier model on an existing architecture if the base model is not available and the instruction trained version doesn't fit with what you want. There are techniques like ablation, but those could have other effects on the model, and there can still be lingering effects of the instruction training in the model that surface less frequently (e.g. on an input not covered by the ablation training).
Otherwise, fine tuning is definitely the way to go. However, you need to be careful not to over-tune the model such that it is only tuned to the data you are training it on.