Got curious, sign up, add money to account, try to use. Can't, it's a labs model. Fine, let's enable labs. Can't, unspecified error. Fine, lets contact customer support as instructed, can't no customer support, just a half-assed FAQ, that seems vibe-coded and searched poorly, totally irrelevant answers coming up for all queries tried. Then it hit me:

If AI makes good customer support, then why does no AI company use theirs to provide customer support?

What a coincidence! I just released OpenATP earlier today. OpenATP is an open-source Python package and CLI for agentic automated theorem provers. It includes support for Leanstral with Mistral’s Vibe harness. The previous production Leanstral model was deprecated on May 22nd. I will update the package to point to Leanstral 1.5 ASAP!

GitHub: https://github.com/henryrobbins/open-atp

Docs: https://open-atp.henryrobbins.com

Tangential, but I'm pretty sad about EU having absolutely nothing in the actual SotA LLM market. Especially given the recent events of US completely restricting the actual SotA models.

Has this been just pure lack of funding and infra?

I'm not sure I understand the Weights policy. This site says the weights are Apache-licensed, suggesting it's open weights. But I can't find a download link. Their Huggingface profile seems to only provide an earlier snapshot [0]. Any pointers on whether/where we can or will be able to download the weights?

[0] https://huggingface.co/mistralai/Leanstral-2603

404?

https://web.archive.org/web/20260630223430/https://docs.mist...

Discussion about Leanstral 1: https://news.ycombinator.com/item?id=47404796

"Page not found" for me. Did you manage to access this? What is this about?

Lean 4 and Idris 2 are underrated, and likely great for LLM's to code in (since they provide additional guarantees)

Interesting that this only specialized for Lean4 and not for similar like Coq

Real talk, does anyone use anything from Mistral because it performs the best, by whatever secular metric of your choosing? Or is it only used "because EU"? Just focus on answering the question. I wonder if anyone has observed it perform better on any objective metric in any rigorous setting.

I am getting 404 right now

Registered due this news. But I must connect to GitHub to use "Code"? That seems limited?

Is this useful for specifying programs too or only theorems?

This went to market horribly (if you can even call it that), just look at the comments. Mistral played themselves big time over the past ~18 months. Non-competitive products and models combined with bad marketing and GTM...Oh Europe

Software in general, AI as well, is a rich get richer market. The big American companies can afford to (and very much do) scoop up European talents and upcoming European companies. And if they don't want to buy them, they can undercut them to bankruptcy. We live in a colony economy, with human capital as the raw produce, and it all gets funneled to the USA.

The only way to avoid this is to stop playing the game as it is today, and start using proper industrial policy to build up a competitive industry (like China did). There has been no appetite for that the last decades, but Trump is making it completely clear that the state is back, and Europe is slowly acknowledging it as well.

Mistral is generally winning the fights it chooses to fight and that's what they need to do.

Instead of looking at what EU's economy could contribute towards a SotA model it's more accurate to look at what France's economy could contribute, then compare that to the US or China. The scale isn't there. Instead what I like to see is what they can accomplish with that lower scale, and it's stuff like Leanstral, Voxtral and other niche products.

Mistral has raised $4B+ which is a decent chunk of change, albeit not in the league of OpenAI/Anthropic/xAI.

The hard part is justifying pure LLM development financially. Models are all very similar. OpenAI justified it originally by being a 'charity' dedicated to pure research (not financial). Anthropic justified it by saying OpenAI didn't care enough about safety and splitting from them (not financial). Elon justified it by saying that AI would be woke and untruthful unless he built Grok (not financial). Google did Gemini because, well, they're where it all started and because AI research was one of the core missions Larry & Sergey gave it when they started it (but then sat on it for financial reasons).

Then there's the Chinese models. It's unclear what their motives are tbh. I've never seen a really great explanation, only hypotheses. But as they're giving them away for free or very underpriced, their motivation doesn't seem to be financial either.

But Mistral is a normal company. It doesn't have rich backers giving it money based on narratives about cosmic destiny, so it needs to justify what it's doing with ROI. So that more or less rules out large scale LLM training.

There's also EU regulation to consider. When I looked at this in the past I found lots of odd rules that kill off any chance of having a European tech industry. The UK had one that said you could only crawl the internet for research purposes!

https://knowledgerights21.org/news-story/the-uks-copyright-l...

And without the First Amendment you're at much greater risk of being prosecuted for things your models say. See how Germany has taken Google to court over things its models put in its search result pages.

So the benefit isn't clear and the legal risks are very high.

> Has this been just pure lack of funding and infra?

For the most part, yes.

France and Germany are the two biggest EU economies. France has well, Mistral, and we here have a government-funded VC entity that is way too proud [1] to be able to offer a whopping… €125 million (<$150 M USD) for helping European researchers achieve new SOTA in sovereign models. And that sum is not even going to a single challenge winner, it'll be split up among multiple recipients. Don't get me wrong, this is a cool first step, or rather, would have been one about three to four years ago.

It's a pity, really.

[1] (in German) https://www.sprind.org/worte/magazin/verkuendung-next-fronti...

One could make the case that being SotA in 2026 is very costly and not that important for being SotA in 2030 if much more efficient models indeed happens.

The EU simply doesn't have a proper common market, especially when it comes to capital. Having more people than the US and a big economy in aggregate doesn't matter much if you can't efficiently pool resources. Could we in Europe have 100 billion fundraises for a new lab? If not, then it's over and you can give up.

How much does supporting multiple languages as first-class citizens (versus adding a translation layer) cost a model?

Culture

I use their Voxtral Mini STT audio model to automatically transcribe my podcasts into markdown. Out of all the STT models I've tried, it's both the best performing and one of the cheapest! It's really accurate, feeding the episode notes and the podcast description ensures all names are properly spelled, and speaker diarization works really great. (I just do a Gemini flash pass at the end to identify the speakers, so it shows the host name instead of "Speaker 1")

For writing and languange learning it's very decent, especially Mistral Large. The pricing is very good too. I really like the consistently low time to first token and good token per second. Claude, especially in the past, would be very inconsistent, often with outages. Mistral mostly just always works and is very fast.

Technical questions are unfortunately hit or miss. I'm lately pretty much always using a system prompt that emphasizes short answers [1], and Opus regularly one-shots it while Mistral needs a follow up. I use big-AGI as a model router [2] (dumb name, great software), which makes switching midway very easy though. For coding I'm still using Claude Code mostly out of inertia (although I really want to move to an OSS harness) and the one time I tried their `vibe` tool months ago it was a bit rough.

Mistral TTS with diarization is also great and cheap. That's the only thing for which I use their web UI.

[1] Give a short but helpful answer to the question the user asks. When helping with a computer-related task, unless the user asks, don't give any installation or setup instructions, but just get straight to the point. When the user asks a follow up question, give a more complete and longer answer while still not overexplaining. When the user prefaces the question with "short mode off" in any question, give a full and well considered reply.

[2] https://github.com/enricoros/big-AGI

We are not Mistral's target audience. For instance I don't know if Leanstral performs the best as a "formal proof engineering model optimised for automated theorem proving and autoformalization" because I don't even know wth that is or who else does it.

Mistral themselves focus more on b2b; financial services, manufacturing, stuff like that, and they get some big clients that way.

Despite not being their target, I started using them because they have many open models. I continue using them because, yeah EU, but also because the community is great and the tool makes me think more than Claude does. Last, I stick with them because they are one of the few AI companies that are up-front about their environmental impact and are actually trying to minimize it while still providing a decent product.

OCR is off the charts good on every metric you can think of.

LLMs are a near-afterthought at this point if you don’t have data residency requirements. I love them and they’re slightly underrated, their models are consistently well-trained, open, but as you note, behind. There is no metric that will say they’re ahead in anything.

Mistral medium is considerably better at writing than Opus.

I’ve also found it very good at pulling info from pdfs. Even a complicated festival with multiple venues and timetables.

A few months ago, I had some data cleaning to do; their small model was surprisingly efficient and got the job done for 0.2x what I expected to run (Anthropic Sonnet / Haiku). Their TTS / STT is also roughly at the frontier, at least for French.

But I admit I only consider them because they're from France. Haven't seen a dimension where they're competitive for general users

> Mistral because it performs the best, by whatever secular metric of your choosing?

I am. I use them primarily through their vibe CLI.

Reason is simple: They are cheaper (by almost one order of magnitude compared to Claude) and still do the job pretty well.

For small programming tasks, quick prototyping, refactoring or anything verbose and not requiring a context too large: I first go to Mistral and then eventually to Claude if I'm unsatisfied.

I also found out some of their models to be more responsive than OpenAI ones (which is not so surprising considering the size).

My tasks are mainly C++ and Python programming. People in other languages might not share my enthusiasm.

I use it as my workhorse for coding and general chat questions, because it's good enough 80% of the time, and indeed it's french/european (with heavy US capital tho...).

We complain too much about not having enough major competitors in the IT space, to not support a burgeoning one even if it's less powerful than SOTA labs

>Just focus on answering the question.

Are you trying to instruct me like an LLM?

For a defense project we're working on, we basically have a hard requirement to use european cloud provider + european llm

We cannot use open source LLMs on-prem, I asked. So that's basically a hard requirement to use mistral, even though Chinese models are strictly better on every dimension.

I like the models for creative writing. They have a distinct voice that is different from the other llms.

I use it because it’s a simple, convenient and cheap OCR api. Specifically via my ringbinder[1] tool.

[1]: https://github.com/maxim/ringbinder

I use it because EU and API pricing is decent to me. And support is awesome also. They reply the same day or at most the next day, and they follow the ticket great. It isn't that bad, but neither the best.

I still prefer Mistral Nemo 12B for text summarisation tasks. It has a nice style. The Mistral Small 24B is also decent. I have a YouTube transcript summariser which I like these for.

However these days I usually have Qwen 3.6 27B already loaded so I mostly just use that instead.

I used them because they had the fastest chat response. (Dont think that’s the case anymore, and they introduced some UI blocking feature on load which is irritating, but still use it mainly due to habit)

I liked that their website didn’t ask for my phone number, IIRC.

just used mistral for a database/scraping creation tool and ended at <10k€ in token costs (via openrouter), beating gpt5.4-mini in output quality and speed and costs after actual testing A/B fairly. so its a super scoped task to be performed hundreds of thousands of time for some automation and mistral just did it better across all dimensions that gpt-5.4-mini. of course thats not a headline in terms of frontier model competitiveness, but for "the boring parts" it just was flat out better than anything else consistently. bonus points it handles mixed-language-content with nuances surprisingly well to turn web content in the wild into structured data really good and fast.

I use Claude with Kiro at work and at home Mistral stuff with pro subscription for coding and research. I don't see a fundamental difference between them in results, at least with what I use them for. The cost is lower too, but it is not the only reason why I won't use Claude at home.

Vibe (the CLI Kiro-like tool) feels more "Claude-ish" as it has more terse answers, while the Le Chat can get quite chatty if you really push it ( * ). Prompting it right gets Le Chat more focused and back to being terse.

( * ) Le Chat is by default in "Fast" mode. It also has "Think" and "Research", but I've not felt the need to use those yet.

As a practical example:

I'm doing some on-and-off family research when I can and have some time since it's the rabbit-holest of all rabbit-holes.

I already knew the answers, but decided to test how well Le Chat does. I fed a 208 year old church book page to Le Chat and asked if it can read that handwritten old cursive Swedish and tell me the contents. Le Chat had a look and then explained it could not, but it pointed me to something I had completely missed before, despite having looked for it: The Swedish Lion OCR model by Riksarkivet (Swedish National Archives) which is purpose-built for OCRing such records (the tool by READ-COOP including The Swedish Lion is at http://www.transkribus.org/).

Then, Transkribus confirmed my existing information and confirmed that I can kind of sort of read that cursive too. I did not expect any new findings here, just to verify against known facts and things were OK. The tip to Transkribus was very nice and I'll be using that tool more.

After this, I asked if Le Chat could dig up some more information about the person (it's a Swedish-Russian noble family connection so a lot of written material exists). Le Chat went about it, summarized real web pages, and the information matched with what I knew already, which was good, no hallucinations or such. I prompted Le Chat further towards the parents and grandparents and so on, it kept on replying with factual summaries and references to web pages that actually exist.

This was all in all very good. For what I earlier spent perhaps a weekend or two (taking like two weeks of wall clock time) in total, I could dig up basically the same information, with sources referenced, in a fraction of the time. Even if Le Chat could not read the page, it pointed to Transkribus and that was very helpful.

The point is: Mistral performs well for me and I see no reason to use something else. There's also the option to turn off the "use my inputs to learn the model". I don't know if Claude has it or not.

Edit: formatting Edit: READ-COOP SCE does the Transkribus