Mistral AI releases Mistral Large 4 weights

Mistral AI releases Mistral Large 4 weights — 1-trillion parameter flagship

Mistral AI has published the weights for its flagship multimodal MoE model, Mistral Large 4 (Le Chonk), on Hugging Face. The model features 1 trillion parameters (52B active) and a 1M context window, and was trained in Europe under a modified MIT license.

Compiled by Sergey KostenchukPublished 2026-10-29Updated 2026-10-29

2026-10-29 Research NVIDIA

🤖 Mistral releases Mistral Large 4 weights (1 trillion parameters)

Mistral AI has uploaded the weights for its flagship Mistral Large 4 (Le Chonk) to Hugging Face: a multimodal MoE with 1 trillion parameters (52B active), 128 experts + 1 shared, and a 1M context window. It was trained from scratch on 3,800 NVIDIA Grace Blackwell units in Europe and fine-tuned via RL for code and cybersecurity: 48% on the Coding Agent Index, 51% on the Cyber Index. API pricing is $0.70/$2.10 per 1M input/output tokens.

🌍 The largest open model trained entirely in Europe: companies and government agencies can inspect, fine-tune, and deploy it on their own infrastructure without relying on a third-party API. The license is a modified MIT, but Model-as-a-Service requires Mistral's permission.

👤 No installation required — Mistral API (mistral-large-4), OpenRouter, or NVIDIA NIM. Low/Medium/High modes offer a balance of quality and price.

Source 1: https://mistral.ai/news/mistral-large-4-oss/ Source 2: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B

Sources