Mistral’s Le Chonk leads open coding, trails China overall

By Rebecca Stone

Artificial Analysis gave Paris-based Mistral's Large 4 a 26-point score on version 5 of its Intelligence Index, below DeepSeek V4.1 Flash and GLM-5.3 at 32 points each. Mistral's largest model yet finished ahead of the 23-point median for similarly sized open models, but it did not catch China's leading systems.

That 26-point result came in the first independent test of the final model.

Mistral released Large 4's weights on Hugging Face three weeks after its public preview. Mistral describes the model, nicknamed “Le Chonk,” as accepting text and images while producing text and code. It has about 1 trillion total parameters, with 52 billion active at each step.

According to SGLang inference documentation, the standard version runs on eight Nvidia B300 GPUs. A heavily compressed variant runs on four GB300 GPUs. Mistral says self-hosting requires substantial hardware.

The combination of open weights, European training and a commercial-cloud license shapes Mistral's pitch to organizations seeking local inference. It also puts Large 4's coding and cost advantages against Chinese models that remain stronger overall.

What did Mistral change after the preview?

Mistral continued reinforcement-learning training after the preview, focusing on coding and cyber defense. Researchers Pierre Stock and Corentin Petit said the team built tasks in-house and made them harder as training progressed.

On the press call, Stock and Petit said: “The only limit on performance is how difficult a task the team can still create and verify.” Mistral calls the effort a “two-year bet” that is now paying off.

Does Large 4 lead open models on coding?

• After three more weeks of reinforcement learning, Artificial Analysis recorded a rise in Large 4's Terminal-Bench 4.0 score from 28 percent to 41 percent. GLM-5.3 scored 40 percent.
• Mistral says Large 4 reaches 48 percent on Artificial Analysis' Coding Agent Index, a practical tie with GLM-5.3. Mistral says it uses 47 percent fewer tokens to get there. The index places Claude Code with Claude Opus 5.5 at the top with 62 points.
• On Cognition's new FrontierCode 2.0 benchmark, Large 4 leads the open models at 19.7 percent. Claude Opus 5.5 scores 46 percent.
• In blind comparisons run by Surge, testers preferred Large 4's code over code from GLM-5.3, Kimi K3 and DeepSeek more often than not. Against Opus 5.5, Large 4 won 22 percent of matchups.

Mistral's published comparisons omit Moonshot's Kimi K3.1 Pro because its weights are not public. Moonshot plans to release them by November 4, 2026. Kimi K3.1 Pro leads Terminal-Bench 4.0 at 67 percent, compared with Large 4's 41 percent.

Why does Mistral call Large 4 the leading cyber model?

Mistral calls Large 4 the leading open model for cyber defense. The company says the final version scores 51 percent on Artificial Analysis' Cyber Index, which tests whether a model can find, reproduce and patch security holes without breaking existing features.

On Mistral's own chart, DeepSeek V4.1 Flash leads the open models at 57 percent, followed by MiMo-V2.6-Pro at 56 percent. Large 4 ranks third, ahead of GLM-5.3 Flash at 50 percent. Kimi K3.1 Pro leads the entire index at 64 percent.

The extra training barely changed the cyber result. Artificial Analysis says the preview scored 50 points, while the final version gained one point.

Mistral also says Large 4 rejects 96.6 percent of malicious cyber requests on average, more than any other open model. The model must still handle legitimate defensive tasks.

Mistral acknowledges that the model could make real attacks easier. It plans to tighten safeguards in the coming months.

Orange Cyberdefense is testing a special version for penetration testing. Access requires a review and contractual restrictions.

Mistral also reports progress on office work. The company says Large 4 trails only Kimi K3 among open models at editing financial spreadsheets and searching the web.

What do the API price and license change?

Mistral prices Large 4's API at $0.70 per million input tokens and $2.10 per million output tokens.

The preview list price was $1.36 per million input tokens and $4.18 per million output tokens. Mistral discounted those rates to $0.68 and $2.09 for two weeks at launch, so the released pricing roughly matches the old promotion.

For comparison, Mistral lists Kimi K3.1 Pro at $2 per million input tokens and $10 per million output tokens.

Mistral released the weights under a modified MIT license. Use, modification and commercial deployment are free. A company offering Large 4 as a commercial cloud service needs a separate license from Mistral.

Mistral mainly pitches the model as sovereign AI. The company says it trained the model entirely in Europe and that it can run on a customer's own servers. Mistral also made it available through Microsoft Foundry and other platforms.

That split gives self-hosting buyers a permissive route while requiring commercial cloud providers to negotiate with Mistral. It is a central distinction in the model's pitch to organizations seeking local inference.

How far behind is Large 4 overall?

Artificial Analysis released version 5 of its Intelligence Index on Wednesday. It then rated the final Large 4 in high mode at 26 points. Three-fifths of the index now uses tests whose tasks or solutions are not public; the nine tests include Terminal-Bench 4.0 and FrontierCode 2.0.

The preview scored 25 points, so three more weeks of training added one point overall. The gap with GLM-5.3 is not new: in the old index, Large 4's preview scored 38 points and GLM-5.3 scored 45.

Large 4 generates 161 tokens per second, up from 111 for the preview. In Artificial Analysis' head-to-head test, only DeepSeek V4.1 Flash and Muse Wonder are faster.

An index task costs $2.06 on average with Large 4, down from $5.54 for the preview. That is cheaper than GLM-5.3 at $5.01 and Kimi K3.1 Pro at $3.68, but dearer than DeepSeek V4.1 Flash at $0.60 and MiMo-V2.6-Pro at $0.26.

Artificial Analysis still places Large 4's per-token price on the high side for open models of similar size.

Mistral also promises token efficiency, and the full index now provides a check. Large 4 generated 110 million output tokens, down from 150 million for the preview. The median for comparable models is 100 million, leaving Large 4 slightly above it.

Google's Gemini 4 Carbon leads the overall index with 59 points, one point ahead of Anthropic's Claude Opus 5.5. So far, only select cybersecurity partners can use Gemini 4 Carbon.

Kimi K3.1 Pro is the highest-ranked open model at 46 points, but its weights have not been released.

What happens next?

Mistral's team said on the press call that it works with a fraction of the resources available to US and Chinese labs. The team said the model would be at the top with 10 times more capital.

Large 4 is intended as a base for specialized Mistral models. Mistral did not comment on whether it will release smaller versions built from it.

With safeguards due for tightening in the coming months and Kimi K3.1 Pro weights due by November 4, 2026, Mistral's coding lead and cyber ranking face their next direct tests.