Mistral releases Mistral Large 4 for free, matching GPT-6.1 Terra's coding performance

Summary

Mistral released the 1-trillion-parameter MoE model 'Mistral Large 4 (Le Chonk)' as open weights on October 29, 2026. The company claims the model's coding agent performance is on par with GPT-6.1 Terra and that it outperforms various Chinese open models in human-evaluated coding benchmarks.

Why it matters

The free availability of a model with 1 trillion parameters and 1 million token context capacity pushes the boundaries of open-weight models. This puts pressure on closed labs in terms of price and capability. It also intensifies the US–EU vs. China open-model race. Mistral explicitly positions the release as a move against DeepSeek, GLM, and Kimi. For developers, this lowers the barrier to deploying frontier-class coding agents without per-token API costs.

Highlights

• Total 1 trillion / active 52 billion MoE parameters, 1 million token context, released under the proprietary 'Mistral Ecosystem Open-Weight License' which still requires a commercial license for MaaS providers
• Mistral Large 4 + Mistral Vibe scored 48 points on the Artificial Analysis Coding Agent Index, tying with GPT-6.1 Terra + Codex. However, on the General Intelligence Index v5.0, it scored 26 points, trailing DeepSeek V4.1 Flash and GLM-5.3
• Human-evaluated coding win rate is above 50% against GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash. The model is evaluated alongside Claude Opus 5.5 in comparisons

Critical take

The 'on par with GPT-6.1 Terra' headline relies on a single agent-pairing benchmark (model + Mistral Vibe vs. model + Codex). In the broader Intelligence Index, the model trails the Chinese competitors Mistral claims to dominate. Therefore, the equivalence claim is narrow in scope, context-dependent, and consciously ignores these lower general scores. Additionally, the 'open-weight' label is softened by a license that still requires commercial terms for Model-as-a-Service monetization, making the model less open than Llama or Qwen.