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.
Mistral isn't even close to the most used AI services in the EU, there are very few industries where EU sourced/hosted models are required (mostly ones where the US also either mandates US models or is seeking to...) and the potential behaviour of the US government (rational and irrational) is at least as big a contributor to European companies being cautious about dependency on US cloud models as EU regulation. Regulatory hurdles exist to building datacentres in the US and basically everywhere else, and don't favour Mistral over any other company, and European states and the EU are busy funding [notional] competitors
It is not, even really stretching the definition, a state sanctioned monopoly.
France is the last country to lack electricity, the nuclear parc is underexploited despite being World's first exporter with around 8 EPR by year last time I checked.
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.
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 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.
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.
No. I've had pretty much every model I've worked with (list too long to recite) spit out the odd Chinese, Korean, Thai or other Asian chunk of letters from time to time. ChatGPT in various incarnations has suddenly switched to Chinese in mid-output and remained that way until told to switch back. The only one I haven't really seen this from as far as I can recall is Gemini, but then again, I haven't used it as much because it was performing quite poor for what I was doing.
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.
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.
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.
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 ;-)
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.
Yeah ... I, erm... can't take that blog post remotely seriously.
Engineering a situation and calling it a vulnerability ?
What a joke. I'm sure you could find a way of engineering a fake vulnerability with Anthropic or OpenAI too and then write a fancy blog post about it.
AND ANYWAY ... whether you are using US or Chinese models, what stupid idiot does not review the slop written by LLMs anyway ? Especially if you are writing for a security-sensitive environment ?
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.
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.
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).
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.
reply