Mistral releases Large 4, a coding model that trails Chinese rivals

Mistral has published the weights for Large 4, a model nicknamed “Le Chonk” that contains roughly one trillion parameters but only activates 52 billion per step. The release comes with a modified MIT license, allowing independent use and deployment on private hardware provided you do not resell the service as a cloud offering.

While the model performs competitively on coding tasks, independent benchmarks show it falls behind leading systems from China and lags significantly behind closed models from Anthropic and Google.

The company claims the system excels in speed, generating 161 tokens per second, and maintains a high refusal rate of 96.6 percent for harmful cyber requests.

Training behind the scenes

Mistral released the weights three weeks after a public preview. The model accepts text and images as input and produces text and code outputs. Running the standard version requires eight Nvidia B300 GPUs, while a compressed variant fits on four GB300 GPUs.

Researchers Pierre Stock and Corentin Petit explained that the team continued training the model using reinforcement learning. They built internal tasks that increased in difficulty as training progressed. Mistral describes this as a two-year investment that is now yielding results.

The final version improved coding scores from 28 percent to 41 percent on Terminal-Bench 4.0. On Artificial Analysis’ Coding Agent Index, Large 4 reaches 48 percent, which is nearly tied with GLM-5.3 but uses 47 percent fewer tokens to achieve the result.

In blind comparisons by Surge, testers preferred Large 4 over code from GLM-5.3, Kimi K3, and DeepSeek more often than not. However, Large 4 won only 22 percent of matchups against Claude Opus 5.5.

The company notes that comparisons exclude Moonshot’s Kimi K3.1 Pro because its weights are not yet public. Moonshot plans to release them by November 4, 2026. Kimi K3.1 Pro currently tops Terminal-Bench 4.0 at 67 percent.

Cyber defense ranking

Mistral states that Large 4 is the leading open model for cyber defense, scoring 51 percent on Artificial Analysis’ Cyber Index. This metric measures the ability to find and patch security holes without breaking existing features.

However, Mistral’s own charts show two other open models ahead of it. DeepSeek V4.1 Flash scores 57 percent and MiMo-V2.6-Pro scores 56 percent. This places Large 4 in third among open models, just ahead of GLM-5.3 Flash at 50 percent. Kimi K3.1 Pro leads the entire index at 64 percent.

Additional training barely moved the score here. The preview version scored 50 points, and the final version is only one point higher.

The model does better at refusing harmful requests. Mistral says it rejects 96.6 percent of malicious cyber requests on average, which is higher than any other open model. It remains capable of handling legitimate defensive tasks.

Mistral admits the model could make real attacks easier and plans to tighten safeguards in the coming months. Partners like Orange Cyberdefense are already testing a special version for penetration testing. Access requires a review and comes with contractual restrictions.

Price and licensing

Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens via the API. The list price during the preview was $1.36 and $4.18. Mistral cut the price to $0.68 and $2.09 for two weeks at launch, effectively matching the old discount.

For comparison, Kimi K3.1 Pro costs $2 for input and $10 for output.

Mistral released the model under a modified MIT license. Use, modification, and commercial deployment are free for direct use. Companies that offer Large 4 as a commercial cloud service need a separate license from Mistral.

The company pitches the model as sovereign AI because it was trained entirely in Europe and can run on your own servers. If you do not wish to host it yourself, it is available on Microsoft Foundry and other platforms.

Independent test results

Artificial Analysis released version 5 of its Intelligence Index on Wednesday. It has since rated the final version of Large 4 in high mode, which scores 26 points. Three-fifths of the index relies on tests whose tasks or solutions are not public.

The preview scored 25 points, so the three extra weeks of training added just one point overall. The index still falls short of the strongest open models from China. DeepSeek V4.1 Flash and GLM-5.3 each score 32 points, and MiMo-V2.6-Pro gets 30.

Large 4’s strengths are speed and cost. The model generates 161 tokens per second, up from 111 for the preview. In Artificial Analysis’ head-to-head comparison, 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. This makes Large 4 cheaper than GLM-5.3 ($5.01) and Kimi K3.1 Pro ($3.68), but much more expensive than DeepSeek V4.1 Flash ($0.60) and MiMo-V2.6-Pro ($0.26).

Mistral also claims efficiency with tokens. The model generated 110 million output tokens across the full index, compared with 150 million for the preview. The median for comparable models is 100 million, so Large 4 sits slightly above that average.

The gap at the top of the index is wide. Google’s Gemini 4 Carbon leads with 59 points, one point ahead of Claude Opus 5.5. So far, though, only select cybersecurity partners can use Gemini 4 Carbon.

The top open model is Kimi K3.1 Pro at 46 points, but its weights have not been released yet.

Mistral believes it is on track. The team said on the press call that it works with a fraction of the resources US and Chinese labs have and would be at the top with ten times more capital. Large 4 is meant to serve as the base for specialized Mistral models. The company would not comment yet on smaller versions built on it.