Mistral AI Releases Mistral Large 4 'Le Chonk' as Public Preview, Weights Due End of October

The model activates 49 billion parameters per token, roughly 4.7 percent of its total, and pairs that routing with a 1.6-billion-parameter vision encoder and a 1-million-token context window, per the model documentation. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters, using data that spanned more than 160 languages, including every official language of the European Union. The company has not published the expert count, top-k routing or layer layout, saying those details will ship with the weights. No license has been announced.

Mistral's strongest reported results are in cybersecurity. The company reports 93 percent on Cybench and 82 percent on CyberGym-E2E, which it says places Mistral Large 4 in the global top five on the Artificial Analysis Cyber Index. Mistral also makes a structural argument about the CyberGym-E2E score: it states that several closed frontier models score near zero on that benchmark because they refuse the task outright. Reproducing a vulnerability to confirm it is real is standard defensive work, and provider-level refusals block it.

In agentic coding, Mistral reports 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA and 28.3 percent on Terminal-Bench 4.0, for a combined Artificial Analysis Coding Agent Index of 49.8 percent. Mistral notes that these runs were evaluated privately ahead of the harness going public, so they are not yet independently reproducible.

A blind human evaluation run with Surge AI offers a separate signal. Professional annotators rated Mistral Large 4 Preview 3.74 out of 5, second among five models, ahead of GLM-5.3 at 3.60 and Kimi K3 at 3.59, but behind Claude Opus 5 at 4.22. On safety, the model resists 93.3 percent of attacks on Lakera's B3 benchmark and scores 1.691 out of a maximum 2.0 on KORABench.

Against open-weight rivals, Mistral Large 4 sits below DeepSeek V4 Pro and Kimi K3 in total parameter count. DeepSeek V4 Pro carries 1.6 trillion total parameters with 49 billion active, a 1-million-token context window, no native image input, MIT licensing and weights already on Hugging Face after an August 2026 release. Kimi K3 lists 2.8 trillion total parameters and roughly 104 billion active, native image input, a 1-million-token context window, modified MIT licensing and weights available since 27 July 2026, priced at $3.00 per million input and $15.00 per million output tokens. GLM-5.3 does not officially publish its parameter counts or active-parameter figure, has no native image input, offers a 1-million-token context window under its own license, and costs $1.40 per million input and $4.40 per million output tokens after its 14 August 2026 release.

The preview API supports function calling, structured outputs, document question answering, batching, and the Agents and Conversations endpoints. Cached input is priced at $0.14 per million tokens, which changes the cost profile of long-context agent loops at a 1-million-token window. Mistral notes that the full 1.05 trillion parameters still have to sit in memory, so the active-parameter count sets compute requirements rather than the hardware bill. Self-hosting is not possible until the weights are released.