Mistral opens Large 4 weights: strong at code, but Chinese models score 32 to its 26

French company Mistral has published on Hugging Face the weights of its largest model, Mistral Large 4, which it calls "Le Chonk". It is an open model that accepts text and images and outputs text and code. It has about a trillion parameters, but only 52 billion are active at each generation step; the rest "sleep", a consequence of the mixture-of-experts architecture.

The independent lab Artificial Analysis combines the results of nine tests into a single intelligence index. Large 4 scored 26 there. That is above the median for open models of similar size (23), but below China's DeepSeek V4.

1 Flash and GLM-5.3, which score 32, and MiMo-V2.6-Pro, which scores 30.

It is far from the market leaders: Google's Gemini 4 Carbon scores 59 and Anthropic's Claude Opus 5.5 scores 58.

Code: the leader among open models, but not among all models

After the public preview released three weeks ago, Mistral kept fine-tuning the model with reinforcement learning. The model receives tasks that the team devised itself and can verify automatically, and it is rewarded for correct solutions. The difficulty of the tasks increased as training went on. Researchers Pierre Stock and Corentin Petit told journalists that the quality ceiling is set only by how hard a verifiable task the team is still able to come up with. Mistral itself calls this a "two-year bet".

Here is what came out of it:

• Terminal-Bench 4.0 (an agent solves tasks in a terminal): up from 28% to 41%. GLM-5.3 scores 40%. • Artificial Analysis's Coding Agent Index: 48 points, practically on par with GLM-5.3 but using 47% fewer tokens. Claude Code running on Opus 5.5 scores 62. • Cognition's FrontierCode 2.0: 19.7%, the best result among open models. Opus 5.5 scores 46%. • Surge's blind comparison (testers pick the best code without knowing who wrote it): Large 4 beats the code of GLM-5.3, Kimi K3, and DeepSeek more often than not, but wins against Opus 5.5 only 22% of the time.

A caveat about Moonshot's Kimi K3.1 Pro: it does not appear in Mistral's comparisons because its weights are still closed. Moonshot promises to open them before 4 November 2026. Kimi K3.1 Pro already scores 67% on Terminal-Bench 4.0, so Large 4's lead in open-weight code will most likely not last long.

Cybersecurity: a claim to first place, a real third

Mistral calls Large 4 the best open model for cyber defense. The Cyber Index tests whether a model can find a vulnerability in a codebase, reproduce it, and fix it without breaking anything. Large 4 scores 51 there. But on Mistral's own chart, it trails DeepSeek V4.1 Flash (57) and MiMo-V2.6-Pro (56), which makes it third among open models. Across all models, Kimi K3.1 Pro leads with 64. The preview scored 50, so the extra three weeks of training added one point here.

It is also expensive: $1.21 per task, while DeepSeek V4.1 Flash scores higher for a small fraction of that. On the plus side, the model refuses 96.6% of malicious requests for cyberattacks, which Mistral says is better than any other open model. The company acknowledges that the model could make real attacks easier and promises to tighten safeguards in the coming months. A separate pentesting version is already being tried by selected partners such as Orange Cyberdefense, with vetted access and contractual restrictions.

Speed, price, and a license with a catch

Large 4 looks strongest on cost and speed:

• 161 tokens per second versus 111 for the preview; only DeepSeek V4.1 Flash and Muse Wonder are faster. • The average index task costs $2.06 ($5.54 for the preview). GLM-5.3 costs $5.01 and Kimi K3.1 Pro $3.68, but DeepSeek V4.1 Flash costs just $0.60 and MiMo-V2.6-Pro $0.26. • Across the whole index, the model generated 110 million output tokens, compared with 150 million for the preview. The median for comparable models is 100 million. • API price: $0.70 per million input tokens and $2.10 per million output tokens. During the preview it was $1.36 and $4.18. For comparison, Kimi K3.1 Pro costs $2 and $10.

The license is a modified MIT. You can use, modify, and commercialize the model for free. But companies that want to sell Large 4 as a cloud service need a separate agreement with Mistral. Running the model yourself is not easy: according to the SGLang documentation, the standard version needs eight Nvidia B300 GPUs, and the compressed one needs four GB300s. If you don't have the hardware, the model is available in Microsoft Foundry, among other places.

What this means in practice

Mistral sells Large 4 as "sovereign" AI: trained entirely in Europe and able to run on a customer's own servers. For European companies with strict data requirements, that is a selling point. For tasks where price and speed matter more than absolute peak quality, the model looks like a reasonable option. Those who need the best results in complex code and security should for now look at Claude and Gemini, and among open models at the Chinese ones.

The methodology deserves a note as well. Version 5 of the index came out on Wednesday, and three-fifths of its tests are closed: the tasks or solutions are not published. On the old index, the gap was the same: the preview scored 38 against GLM-5.3's 45. Mistral says it works with a fraction of the resources of American and Chinese labs and that, with ten times the capital, it would be at the top. Large 4 will become the basis for the company's specialized models, but Mistral has so far said nothing about smaller versions built on it.