Seems like it's multimodal and visual intelligence capabilities are better than Deepseek 4.1 Flash, even though coding is not equal to DS 4.1 1017. The cost is about 10x of Deepseek though comparing on OpenRouter.
It's gonna be interesting to see the final version, since this is a preview model still. I'm not completely sure as to how this works, but from what I gathered it will be further trained during this period, so it might improve a bit.
Since cost is so much higher than deepseek though it really seems like the biggest reason to use it would be "in the mistral subscription" for personal use or "because you need to abide by european data privacy laws" for companies.
Mostly I'm happy they're still in competition, as I mentioned already elsewhere.
Are they still in competition? Or are they a government sanctioned monopoly? I’m not sure that will be competitive. There are no market forces to do so but pride.
Just the lack of capital, lack of electricity, lack of datacenters, lack of gpus. Of course to navigate all that you have to fight endless regulations in the EU.
No one prevented other people from running their own AT&T phone company in the US. Except the lack of capital, and endless regulations with running telephone polls across the US.
Google even tried more recently to run their own fiber optics and gave up in the end because of entrenched regulations making it impossible to challenge the modern monopolies internet providers have.
Lack of capital, datacentres and GPUs being the barrier to competition in European AI giants is the exact opposite of a "government sanctioned monopoly"...
That is like saying AT&T didn't have a government sanctioned monopoly, even though the primary thing stopping datacenters in the EU is regulatory hurdles.
A lot of national defense stuff is a de facto "government sanctioned monopoly" and this is certainly the same flavor. Market forces are different if the industry is perceived to be vital to continued sovereignty.
I mean, the model architecture is basically that of GLM 5.3 with slightly more parameters and slightly more active parameters, so it should probably be able to beat GLM 5.3 enough post-training.
Actually, with Trump in the Office, "AI/Data Sovereignty" is quite a strong market forces. From Europe point-of-view, any US company is a potential risk (hello Starlink)
I haven't rigorously tested them, if that's what you mean. N = 10 (more or less) in my case. But I couldn't come close to perfection, except with Astra which cost me around $2 per page.
The weird thing about this model is that it spits out Chinese characters just like the chinese models when chatting with it in German. Happened to me quite often during the past few days.
Opus 4.8 and Fable 5 did the same; they would output random Chinese characters and then say “I made a typo, let me fix it” and it’s just them removing that stay Chinese characters. Haven’t seen this since 5.1 family of models.
Does this suggest Mistral is distilling Deepseek et. al.? Can distillation keep passing through with high fidelity, or is it lossy like copying a VHS tape?
It doesn't, even remotely. It's a typical failure mode in all models, some do this more often. It's not related to samplers either. Not necessarily Chinese characters either, they likely just occupy a large portion of vocabularies. For example Gemini 2.0 used to glitch in Telugu and Ukrainian, Claude 2.x and 3.0 used to switch between random languages for a few words when writing in non-English, once in a blue moon.
I've seen this quite a bit on bad quants as well. I ran an early and very lossy quant of one of the Gemma3 models, and it loved to just stick Chinese and Korean characters into the text, sometimes without seeming to realize it had mixed languages.
Ha, ha. Just like when Tilly Norwood switched to Chinese while interviewing with Piers Morgan [1].
On a serious note, I remember getting spooked when I saw a bunch of Chinese characters dumped out into some of my server logs (pre-LLM era). When I pasted the Chinese content from my log into google translate, it was just complete nonsense. Later I found a bug that was dumping raw binary data into my log file which interpreted as Chinese characters when viewed in a text editor. Perhaps similar bugs existed in the models.
Can you please share:
- what specific input you see this issue with?
- which interface you used (API, playground, specific sdk version or a wrapper app)?
- whether the chinese chars appear in the thinking section or the output?
Sorry it was not one time but it happened easily 5-10x during the past few days on different topics and prompts and I didn't store them. It felt like when pushing the model on some specifics in some niche topics in German (I have been prompting in German and the model responded in German). It was the acutal output and not the thinking traces and it was through the API. I jokingly asked it what the character meant and it didn't even make sense in the context so it seemed like random chinese characters.
Has anything changed? I already ran it on my adversarial programming language benchmark[1] a few weeks ago and it performed pretty poorly (on par with Kimi K3).
Well, Models are hard to just distill down to a single benchmark number, or even a few across multiple different benchmark - they all behave and work slightly differently and you can find some who work or feel better for you and your use-case specifically.
But you're right, Mistral Large 4 seems a much more expensive but overall weaker model across the board.
The reason to use them would be, well, because they're...Mistral. They're in EU, france is large in nuclear energy so the energy cost of training has much less of an environmental impact, the data protection laws are very good in EU. This can really be a deciding factor for many EU companies, and I think it's good that they now have something which is at least semi-usable. (apparently)
It's also good that other options exist, even if they are worse. Deepseek is kind of like a large specter looming over every model release, because it's just SO cheap and good. That doesn't mean we should abandon every other model.
Still, the model is not really competitive on price and abilities alone. I'm mostly happy to see Mistral still in the fight.
I say this with a slight tongue in cheek, but Mistral does have a better option: they recently added GLM 5.3 to their platform (and it works rather well). Not trained in the EU, though, of course ;-)
The organization I work for uses GPT and Claude models hosted in EU. I don't see any practical reason to use models made in Europe. Even if you don't want to work with American big tech there are plenty of companies in the EU serving Chinese open-weight models.
Interestingly Mistral has been selling GLM5.3 inference hosted in Europe on their own hardware, for those who are OK using a "not developed in Europe" model but run on hardware located in and controlled by a European corporation.
> for those who are OK using a "not developed in Europe" model but run on hardware located in and controlled by a European corporation.
For the record this is also possible in Europe on Privatemode[1] which also gives you enclave-based processing. And access to GLM-5.3 Flash as well as GLM-5.3
Indeed, I see no difference between a chinese model hosted in Europe vs a Europe-grown model.
You could argue that potentially you can train a backdoor into a foreign-grown model (that has been demonstrated to be possible in toy examples), but I would highly doubt that this would be the case today, given the potential reputational damage if discovered.
Remember Stuxnet? An american worm that was completely benign, travelled on usb drives in hiding, and only activated when it guessed that it's on Iranian plants, and iirc changed very subtly some very specific parameters of operations.
Here, with a chinese model, it would be even simpler - you finetune a model to activate certain operations when it discovers certain environments, but to behave completely normal otherwise. Kind of like a sleeper agent. And with current tech we wouldn't even be able to spot this.
It's not like they didn't sandbox it or limit the access. There were reasonable measures, but not really serious. They made a kodoku pit and didn't watch it.
The agents started hacking because they were sandboxed.
It was filtered, but had a bug, and multiple frontier agents, givent enough time first figured out how to communicate and coordinate with each other and then how to break out of it.
For some benchmarks agents need some sort of a limited internet access - otherwise the benchmarks would be pointless. Good luck implementing a good filter on the internet - if you have a way to do it reliably please share, I'm sure multiple people here would love to hear it.
> you could argue that potentially you can train a backdoor into a foreign-grown model
If we're wearing tinfoil hats then do we know the NSA have not planted stuff in Anthropic and OpenAI anyway ? ;)
I think the Chinese would be happy enough to see OpenAI and Anthropic suffer because people discover open-weights can get you most of the way there for a fraction of the cost.
You don't have to target the end users. Economic impact on the US providers would be sufficient.
The Chinese are also no doubt seeking to play the long game of hoping to (eventually) rebalance the US-centric world order in an Easterly direction.
I recall seeing something about that. I don't know of anyone still using such an ancient model when much newer Qwen 3.5-120, 3.6-35B-A3B, 3.8-27B and 3.8-flash-next all exist and score much higher on all conceivable metrics. I would like to see if anyone can replicate the same thing on a 'current' Qwen released in the past 6 months.
DS4 hosted by deepseek used through the API doesn't have a problem talking about tiananmen square and Uyghurs, but it thinks it's Claude when it speaks about it. I'm not sure what it all means. That was on v4-pro before the 4.1 switch, so may or may not be reproducable.
Add: I was surprised by the fact that the LLM itself doesn't know it's identity unless the system prompt mentions it.
Maybe not now, but in a few years, we might not need frontier models anymore. Open-weight models are only half a year behind the frontier, and both categories are advancing really fast.
Gemini 2.0 Flash, released just 1.5 years ago, could not even repeat a few paragraphs unchanged. Now, we have Mistral, DeepSeek, and MiMo.
Even as someone happy of seeing Mistral still trying to compete, I would wait at least for the final release and weights before talking again about it, to have a definitive opinion on it.
Previous discussion: https://news.ycombinator.com/item?id=49977979 (2044 points 15 days ago, 1254 comments)
I was confused and thought it's a new model... but no, it's the same. What's the purpose of reposting this as a new story? Am I missing something?
announcement of preview release → preview release → release → final release → final release v2
I think that this link is more helpful: https://mistral.ai/news/mistral-large-4/
Seems like it's multimodal and visual intelligence capabilities are better than Deepseek 4.1 Flash, even though coding is not equal to DS 4.1 1017. The cost is about 10x of Deepseek though comparing on OpenRouter.
Weights will be here https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B at the end of the month.
It's gonna be interesting to see the final version, since this is a preview model still. I'm not completely sure as to how this works, but from what I gathered it will be further trained during this period, so it might improve a bit.
Since cost is so much higher than deepseek though it really seems like the biggest reason to use it would be "in the mistral subscription" for personal use or "because you need to abide by european data privacy laws" for companies.
Mostly I'm happy they're still in competition, as I mentioned already elsewhere.
Are they still in competition? Or are they a government sanctioned monopoly? I’m not sure that will be competitive. There are no market forces to do so but pride.
You are severely misusing the term monopoly.
There is demand for AI that is not US or China controlled. Mistral is serving that demand.
That demand is enshrined in law and being forced through coercion. It is the definition of a government dictated monopoly.
Nobody is preventing other Europeans from setting up AI companies
Just the lack of capital, lack of electricity, lack of datacenters, lack of gpus. Of course to navigate all that you have to fight endless regulations in the EU.
No one prevented other people from running their own AT&T phone company in the US. Except the lack of capital, and endless regulations with running telephone polls across the US.
Google even tried more recently to run their own fiber optics and gave up in the end because of entrenched regulations making it impossible to challenge the modern monopolies internet providers have.
Lack of capital, datacentres and GPUs being the barrier to competition in European AI giants is the exact opposite of a "government sanctioned monopoly"...
That is like saying AT&T didn't have a government sanctioned monopoly, even though the primary thing stopping datacenters in the EU is regulatory hurdles.
"The definition of a government dictated monopoly".
This is a joke right?
A lot of national defense stuff is a de facto "government sanctioned monopoly" and this is certainly the same flavor. Market forces are different if the industry is perceived to be vital to continued sovereignty.
Being the only(?) competitive European model probably makes them a de-facto monopoly for doing sensitive European stuff.
That's rich coming from a country who's top AI contender just announced 80B in yearly losses. Looks a lot like a subsidised duopoly to me.
72B of the 80B was paper losses due to stock exercises and not related to cash flow.
Who's subsidizing it?
Also what about SpaceX, Meta and Google? They are clearly still in the game. They are all ahead of Mistral.
Which government? Sanctioned how? What stops other European companies from competing with Mistral? Is the situation better in the US?
french are pretty proud man. pro-euro europeans too!
I mean, the model architecture is basically that of GLM 5.3 with slightly more parameters and slightly more active parameters, so it should probably be able to beat GLM 5.3 enough post-training.
Actually, with Trump in the Office, "AI/Data Sovereignty" is quite a strong market forces. From Europe point-of-view, any US company is a potential risk (hello Starlink)
;-)
There is one word too many in that last line. The US is no longer a 'potential' risk.
It's well known that every Mistral employee is a fonctionnaire. No.
I'm still waiting for a model that can do sheet music transcription for < $0.10 per page.
Shouldn't be that difficult as there are huge datasets. But somehow, the AI labs cannot do it.
Given how well LLMs do with code, I've been shocked and how poorly they do with sheet music.
Have you set up some sort of benchmark to see if there is actually anything useful or they are complete trash on the task?
I haven't rigorously tested them, if that's what you mean. N = 10 (more or less) in my case. But I couldn't come close to perfection, except with Astra which cost me around $2 per page.
The weird thing about this model is that it spits out Chinese characters just like the chinese models when chatting with it in German. Happened to me quite often during the past few days.
Opus 4.8 and Fable 5 did the same; they would output random Chinese characters and then say “I made a typo, let me fix it” and it’s just them removing that stay Chinese characters. Haven’t seen this since 5.1 family of models.
i’ve seen it and Cyrillic in the 5x family
I saw it in earlier Claude models, but I haven't seen it for a least 7-8 months.
Does this suggest Mistral is distilling Deepseek et. al.? Can distillation keep passing through with high fidelity, or is it lossy like copying a VHS tape?
It doesn't, even remotely. It's a typical failure mode in all models, some do this more often. It's not related to samplers either. Not necessarily Chinese characters either, they likely just occupy a large portion of vocabularies. For example Gemini 2.0 used to glitch in Telugu and Ukrainian, Claude 2.x and 3.0 used to switch between random languages for a few words when writing in non-English, once in a blue moon.
I interpret it to mean there is a lot of Chinese in the training corpus.
I've seen this quite a bit on bad quants as well. I ran an early and very lossy quant of one of the Gemma3 models, and it loved to just stick Chinese and Korean characters into the text, sometimes without seeming to realize it had mixed languages.
Ha, ha. Just like when Tilly Norwood switched to Chinese while interviewing with Piers Morgan [1].
On a serious note, I remember getting spooked when I saw a bunch of Chinese characters dumped out into some of my server logs (pre-LLM era). When I pasted the Chinese content from my log into google translate, it was just complete nonsense. Later I found a bug that was dumping raw binary data into my log file which interpreted as Chinese characters when viewed in a text editor. Perhaps similar bugs existed in the models.
[1] https://www.youtube.com/watch?v=8VOfGcTqqq8
Can you please share: - what specific input you see this issue with? - which interface you used (API, playground, specific sdk version or a wrapper app)? - whether the chinese chars appear in the thinking section or the output?
Thanks.
Sorry it was not one time but it happened easily 5-10x during the past few days on different topics and prompts and I didn't store them. It felt like when pushing the model on some specifics in some niche topics in German (I have been prompting in German and the model responded in German). It was the acutal output and not the thinking traces and it was through the API. I jokingly asked it what the character meant and it didn't even make sense in the context so it seemed like random chinese characters.
happened to me with Gemini too, actually. it sometimes just randomly decides to answer in a different language.
What other languages?
Google also does that surprisingly
you can see the same thing with any models
Havn't seen it for months with other frontier models.
Has anything changed? I already ran it on my adversarial programming language benchmark[1] a few weeks ago and it performed pretty poorly (on par with Kimi K3).
1 - https://bench.killswitch-lang.org/
I see Le Chaton Fat and press upvote, sure. But why? Wasn't it released already 2 weeks ago?
In Vibe CLI app, when you load Mistral Large 4 it swaps out the cat in the header with a larger one
It's a nice little detail
Some details and benchmarks were announced but the weights didn’t get released
Still 9 days before the release of weights apparently.
Le Chaton Fat is a lovely name. I predict the next Michelin restaurant with will have an LLM as the Chef...
compared to Deepseek 1017, I'm not sure why I'd use this?
https://openrouter.ai/compare/mistralai/mistral-large-4-0/de...
Well, Models are hard to just distill down to a single benchmark number, or even a few across multiple different benchmark - they all behave and work slightly differently and you can find some who work or feel better for you and your use-case specifically.
But you're right, Mistral Large 4 seems a much more expensive but overall weaker model across the board.
The reason to use them would be, well, because they're...Mistral. They're in EU, france is large in nuclear energy so the energy cost of training has much less of an environmental impact, the data protection laws are very good in EU. This can really be a deciding factor for many EU companies, and I think it's good that they now have something which is at least semi-usable. (apparently)
It's also good that other options exist, even if they are worse. Deepseek is kind of like a large specter looming over every model release, because it's just SO cheap and good. That doesn't mean we should abandon every other model.
Still, the model is not really competitive on price and abilities alone. I'm mostly happy to see Mistral still in the fight.
I say this with a slight tongue in cheek, but Mistral does have a better option: they recently added GLM 5.3 to their platform (and it works rather well). Not trained in the EU, though, of course ;-)
Yeah, absolutely true. And I am guilty of simply using Mistral as a GLM subscription myself. I just really like the GLM 5.3 model.
I will see where Chonk is better though, since you have both options in one subscription.
Everything feels very European.
Unproductive company that is 3rd place (assuming we are lumping 4 better US models and 3 Chinese models).
The only thing that gives it value is that it's generated in some region in Europe, like random cheese or wine.
But I still can't imagine a usecase outside patriotism.
Privacy can be solved by self hosting a US or Chinese model.
Seems performative.
You are delusional if you think people like a french, italian or swiss cheese just because it was made in said parts of europe.
Or you have no taste.
If you're required to use a European model
The organization I work for uses GPT and Claude models hosted in EU. I don't see any practical reason to use models made in Europe. Even if you don't want to work with American big tech there are plenty of companies in the EU serving Chinese open-weight models.
Interestingly Mistral has been selling GLM5.3 inference hosted in Europe on their own hardware, for those who are OK using a "not developed in Europe" model but run on hardware located in and controlled by a European corporation.
https://docs.mistral.ai/models/zai-glm-5-3
> for those who are OK using a "not developed in Europe" model but run on hardware located in and controlled by a European corporation.
For the record this is also possible in Europe on Privatemode[1] which also gives you enclave-based processing. And access to GLM-5.3 Flash as well as GLM-5.3
[1]https://www.privatemode.ai/
Indeed, I see no difference between a chinese model hosted in Europe vs a Europe-grown model.
You could argue that potentially you can train a backdoor into a foreign-grown model (that has been demonstrated to be possible in toy examples), but I would highly doubt that this would be the case today, given the potential reputational damage if discovered.
Remember Stuxnet? An american worm that was completely benign, travelled on usb drives in hiding, and only activated when it guessed that it's on Iranian plants, and iirc changed very subtly some very specific parameters of operations.
Here, with a chinese model, it would be even simpler - you finetune a model to activate certain operations when it discovers certain environments, but to behave completely normal otherwise. Kind of like a sleeper agent. And with current tech we wouldn't even be able to spot this.
> And with current tech we wouldn't even be able to spot this.
Meanwhile Anthrophic, OpenAI & Co can't even manage to stop their LLMs from hacking into third-party servers until they discover it after the event.
What happened to people's ability to implement basic security architectures like sandboxing etc ?
It's not like they didn't sandbox it or limit the access. There were reasonable measures, but not really serious. They made a kodoku pit and didn't watch it.
The agents started hacking because they were sandboxed.
> The agents started hacking because they were sandboxed.
Not sure your definition of sandbox but mine doesn't include unfiltered internet access ?
It was filtered, but had a bug, and multiple frontier agents, givent enough time first figured out how to communicate and coordinate with each other and then how to break out of it.
For some benchmarks agents need some sort of a limited internet access - otherwise the benchmarks would be pointless. Good luck implementing a good filter on the internet - if you have a way to do it reliably please share, I'm sure multiple people here would love to hear it.
There was a post on here yesterday about embedding bad behavior in models.
>Any open model whose weights have been edited can carry a backdoor, whether it's a task-specific fine-tune, a merged adapter or an abliterated build
https://projectdiscovery.io/research/how-abliterated-models-...
> you could argue that potentially you can train a backdoor into a foreign-grown model
If we're wearing tinfoil hats then do we know the NSA have not planted stuff in Anthropic and OpenAI anyway ? ;)
I think the Chinese would be happy enough to see OpenAI and Anthropic suffer because people discover open-weights can get you most of the way there for a fraction of the cost.
You don't have to target the end users. Economic impact on the US providers would be sufficient.
The Chinese are also no doubt seeking to play the long game of hoping to (eventually) rebalance the US-centric world order in an Easterly direction.
If I was NSA I 'd create obscure websites that are easy to to parse for crawlers and have poisoned instructions there for ALL LLM's to pick.
Kinda like https://github.com/elder-plinius has been doing but covert and with actually harmful (for targets) instructions.
Didn't we get a submission today about Qwen3-Coder producing more bugs when hinted on the user having different ideology than the Chinese government?
Ok I see it's flagged now, probably fake, but the fact that it's possible does make me feel uneasy about both Chinese and US models.
It was so insignificant, it could be equally described by the non-determinism inherent in the models. And that it reads more like a marketing piece.
That's my read anyhow. It's been flagged. The whole thing is kind of rich, considering the Israeli roots of the company.
https://news.ycombinator.com/item?id=50179706
I recall seeing something about that. I don't know of anyone still using such an ancient model when much newer Qwen 3.5-120, 3.6-35B-A3B, 3.8-27B and 3.8-flash-next all exist and score much higher on all conceivable metrics. I would like to see if anyone can replicate the same thing on a 'current' Qwen released in the past 6 months.
> I see no difference between a chinese model hosted in Europe vs a Europe-grown model.
Really? Fascinating. Do you know about Tiananmen Square, the great firewall of China, etc?
DS4 hosted by deepseek used through the API doesn't have a problem talking about tiananmen square and Uyghurs, but it thinks it's Claude when it speaks about it. I'm not sure what it all means. That was on v4-pro before the 4.1 switch, so may or may not be reproducable.
Add: I was surprised by the fact that the LLM itself doesn't know it's identity unless the system prompt mentions it.
> Do you know about Tiananmen Square
People here bringing up the straw clutching argument of "but Tianament Square" is becoming really tiresome.
The fact is a Chinese LLM's lack of knowledge about Chinese politics will not affect 99.99999999999% of users and their workflows.
And for the 0.000000001% who are affected, they don't need an LLM. A quick search of Wikipedia or 5 seconds on Google will suffice.
What the hell has Tiannamen Square got to do with Joe Blogg's vibecoding session ? Absolutely nothing.
Maybe not now, but in a few years, we might not need frontier models anymore. Open-weight models are only half a year behind the frontier, and both categories are advancing really fast.
Gemini 2.0 Flash, released just 1.5 years ago, could not even repeat a few paragraphs unchanged. Now, we have Mistral, DeepSeek, and MiMo.
The TAM is already covered by local models. 10% of a health society could use better models.
We are in the 70s era of gas guzzling before the gas embargo and cheap foreign imports and economy cars reign.
if you care about data-retention policy or you have to use european models ...
Patriotism.
Well when your greatest ally acts like an adversary, who can blame them for the increased sovereign focus?
What's more impressive is the datacenter used for training. It shows you can get "sub-SOTA" for one tenth of the resources of SOTA.
Impressive if these benches hold up to real world use. Not groundbreaking, but certainly in the game.
Would love to see them release some smaller open-weight models.
52B active parameters and my qwen-3.8-flash-next model running at home is supposedly competitive? maybe I gotta test it to believe it. https://artificialanalysis.ai/models/comparisons/mistral-lar...
qwen-3.8-flash-next is suprisingly good, I find it overthinks, but it's quite usable.
"Covered by GDPR" is a very big plus for me. Also very simple training opt-out.
It's quite fast
waiting for pelican bench results (the only benchmark that matters)
NVM. Looks like it already exists:
https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Yeah I did this one two weeks ago when it was first released. Not sure why it's on the HN homepage today.
Because people acclaim European competition on the LLM scene, I guess. We don't have much to show.
Even as someone happy of seeing Mistral still trying to compete, I would wait at least for the final release and weights before talking again about it, to have a definitive opinion on it.