Key takeaways

  • Mistral released the weights of Mistral Large 4, a 1-trillion-parameter multimodal model, on October 29.
  • The API costs $0.70 per million input tokens and $2.10 per million output tokens.
  • Mistral says it ties GLM-5.3 on the Artificial Analysis Coding Agent Index at 48%, but the figures are its own.

Mistral released the weights of Mistral Large 4 on October 29, a 1-trillion-parameter multimodal model that the company nicknames "Le Chonk". Mistral's announcement says it leads open-weight models on coding benchmarks. Through Mistral's API it costs $0.70 per million input tokens.

The weights are on Hugging Face, so anyone can download and run the model, subject to a license with one commercial catch. Mistral had put a preview of the model online earlier this month. This is the final version, with better scores than the preview.

What Mistral released

Mistral Large 4 is a mixture-of-experts model, which means it runs only part of its network for each token. The repository name, Mistral-Large-4-1T-A52B, suggests about 52 billion active parameters out of 1 trillion, but Mistral's post does not state the active count. It takes images as input, and it combines instruct and reasoning modes in one model. Users pick a low, medium or high reasoning level, and high is the default.

Mistral lists self-hosting with vLLM or SGLang. It is also available in Mistral Studio, through OpenRouter, on NVIDIA-hosted endpoints and in Microsoft Foundry. The post does not give a context window.

Price and license

The API costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. That is far below Claude Opus 5.5 at $4 input and $20 output.

The license is a modified MIT license that Mistral calls the Mistral Ecosystem Open-Weight license. Use, modification, fine-tuning and commercial use are allowed. A business that sells the model as a hosted service has to get a separate commercial license from Mistral.

What the benchmarks say

All of these numbers come from Mistral, and some come from tests that outsiders cannot run.

  • Artificial Analysis Coding Agent Index: 48%, which Mistral says is effectively tied with GLM-5.3, the leading open-weight model, while using 47% fewer tokens per task.
  • Terminal-Bench 4.0: 41%, up from 28% for the preview.
  • Cognition FrontierCode 2.0: 19.7%. This is a private benchmark.
  • BrowseComp: 84.5%.
  • Surge Code human preference: Mistral says only Claude Opus 5.5 ranks ahead of it.

Mistral also reports a 96.6% average refusal rate on malicious cyber requests, which it calls the highest among open models. It warns that the model could still help real attackers and says it is changing its guardrails. A cyber-focused variant is limited to vetted enterprise partners.

What we don't know yet

The post has no context window and no Intelligence Index score. Artificial Analysis has not published its own page on the model in the sources we read, so the headline coding number is still Mistral's description of a third-party index. Watch for independent runs on Terminal-Bench and the Artificial Analysis indexes in the coming days. Gemini 4 Carbon now leads the Intelligence Index at 59, for context on where the top closed models stand.

Sources

  1. Le Chonk: Built for Coders, Trusted by DefendersMistral AI
  2. Mistral-Large-4-1T-A52BHugging Face