Mistral Unveils New 'Le Chonk' AI Model It Says Rivals Best Open Systems From China
(Tuesday October 06, 2026 @05:00PM (BeauHD)
from the one-trillion-parameters dept.)
- Reference: 0186051806
- News link: https://news.slashdot.org/story/26/10/06/1614205/mistral-unveils-new-le-chonk-ai-model-it-says-rivals-best-open-systems-from-china
- Source link:
Mistral has [1]unveiled Mistral Large 4, or "le Chonk," a 1-trillion-parameter model that it says is the [2]strongest open-weight AI model developed outside China by a wide margin . The model is entering preview for developers, cybersecurity teams and governments before a broader release later this month. CNBC reports:
> ML4 was trained on 4,000 Nvidia Grace Blackwell GPUs over two months that were deployed in Mistral's own data centers in Europe. When its core parameters are released, ML4 will rank among the top open-weight models globally on aggregate benchmark performance, Mistral said in a statement. The company added that ML4 was the strongest open-weight model developed outside China by a "substantial margin." The new model still lags behind the frontier in areas such as coding.
>
> "The model capabilities will further improve as we scale up our training capacity, following our Series D fundraise," said Guillaume Lample, co-founder and chief scientist of Mistral. "Further, the cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks," he added.
Further reading: [3]Mistral CEO Says US AI Safety Debate Masks Competitors' 'Negligence'
[1] https://mistral.ai/news/mistral-large-4/
[2] https://www.cnbc.com/2026/10/06/mistral-ai-model-le-chonk.html
[3] https://slashdot.org/story/26/09/29/1855248/mistral-ceo-says-us-ai-safety-debate-masks-competitors-negligence
> ML4 was trained on 4,000 Nvidia Grace Blackwell GPUs over two months that were deployed in Mistral's own data centers in Europe. When its core parameters are released, ML4 will rank among the top open-weight models globally on aggregate benchmark performance, Mistral said in a statement. The company added that ML4 was the strongest open-weight model developed outside China by a "substantial margin." The new model still lags behind the frontier in areas such as coding.
>
> "The model capabilities will further improve as we scale up our training capacity, following our Series D fundraise," said Guillaume Lample, co-founder and chief scientist of Mistral. "Further, the cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks," he added.
Further reading: [3]Mistral CEO Says US AI Safety Debate Masks Competitors' 'Negligence'
[1] https://mistral.ai/news/mistral-large-4/
[2] https://www.cnbc.com/2026/10/06/mistral-ai-model-le-chonk.html
[3] https://slashdot.org/story/26/09/29/1855248/mistral-ceo-says-us-ai-safety-debate-masks-competitors-negligence
Meanwhile, back at the benchmarks (Score:2)
https://artificialanalysis.ai/ has them 38 overall on their "intelligence index" compared to mid-50s for the current leading models. That puts it alongside the current leading open weight models, but not ahead at all, or even by a "substantial margin."
1T parameters is huge, although I'm sure active is a much smaller margin of that. If not, the active memory required for inference will be huge. The MoE / MoI system that Qwen-Flash-Next has adopted might be the new path forward here, since it offloads paramet
Re:Meanwhile, back at the benchmarks (Score:4, Interesting)
notice the choose of words:
"The company added that ML4 was the strongest open-weight model developed outside China"
So yes, Chinese open models are still better, they are comparing with OTHER open-weight models outside China (most much smaller).
So competitive with other open-weight models, independent from USA and China, not yet as good for coding, but that should also improve quickly
Valid alternatives is good for competition and as we all know, when the bubble burst, only a few of those will remain. so having alternatives is a good thing
Re: (Score:2)
ohh, and from the documentation: "It features 52B active parameters and 1.05T total parameters, and a 1.6B vision encoder."
So yes, not all parameters are active
Re: (Score:2)
Isn't that the case with all the large models these days?
Re: (Score:2)
That means it's a mixture-of-experts architecture and activates about 5% of its experts for processing each token. Which experts are activated changes per token.
Re: Meanwhile, back at the benchmarks (Score:2)
Whoa, 52B experts really is Le Chonk.
Qwen 3.8 has 6B MoE or a dense 27B published.
Re: (Score:2)
Yeah, that qualifier is only true so long as Meta doesn't open weight one of their newer models and Google doesn't release a new Gemma. I guess I'd care more about "not China" if there was a "not China" quality I was looking for in a model.
Re: Meanwhile, back at the benchmarks (Score:3)
Well, the "not censored" option is pretty much THE quality you get for "not China".
With that said, Qwen is still awesome.
Re: (Score:2)
> Well, the "not censored" option is pretty much THE quality you get for "not China".
> With that said, Qwen is still awesome.
What?!? In my experience, the "western" models are way more censored than the ones from China out of the box, and of course any sane person would want to use abliterated versions of either, anyway.
I have experienced Claude Opus refusing to debug my own source code "because security!", I have experienced Ideogram 4 refusing to render Pizza because of "alignment filters", all while even the non-abliterated versions of Qwen or DeepSeek did no such thing.
I had no problems discussing what happened at Tianamen S
Re: Meanwhile, back at the benchmarks (Score:2)
You can use an abliterated version if you trust the person that did the modification did not also introduce anything else.
Usually, the abliterated versions come from random accounts with names like "fugazi_hentai_xXx". I am not touching that with a 10 foot pole.
Re: (Score:2)
It saw your comment and immediately surrendered.
Re: (Score:2)
So MoE doesn't save on RAM, it saves on compute. You still need to load the model weights into GPU memory otherwise you would cripple the model's speed since each token can be routed to a different "expert" and those weights would need to be copied from system to GPU memory when an new is expert called. Some people do it that way but they are community folks doing it for the LoLs and not anything running at the enterprise level. The trick Qwen-Flash-Next does is move their 51B N-gram embedding table into sy