AI Trend Research
EN

← October 29, 2026 briefing · added in the 00:00 KST update

P0 Open source Open weights Language modelsMultimodal00:00 updateMistral AIHugging FaceNVIDIA

Mistral AI releases open weights for Mistral Large 4 ('Le Chonk'), a 1T-parameter multimodal MoE with 49B active per token, trained in its own European data center

Announced October 29, 2026

Verified (fact)

Mistral AI has released the weights of Mistral Large 4 (nicknamed Le Chonk), the largest model it has built so far. It is a 1-trillion-parameter MoE model with 49B parameters active per token (52B including the embedding and output layers), and it was designed to be multimodal from the start. Mistral said it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data center in Europe. A single set of weights handles instruction following, reasoning, coding and agentic tasks. Hugging Face published an introductory post on release day, and Transformers v5.22.0 supports the model out of the box. The repository is named 'Mistral-Large-4-1T-A52B'. In the community, GGUF quantized versions (including NVFP4) from ggml-org and Unsloth are being shared.

Needs fact-check (claim)

Mistral describes the model as the best open-weight model for coding. It says that although the model is general-purpose, it performs especially well on tasks such as cybersecurity and finance. Comparative benchmark figures are not included in the source material.

Why it matters (interpretation)

A 1T-scale frontier model is now available as open weights, and its training infrastructure is also in Europe, giving a European alternative between Chinese open models and US closed models. With official Transformers support and GGUF quantized versions available right away, it is also easy for companies to try it on their own infrastructure. That said, with 52B active parameters, serving costs are likely to be substantial.

Sources

Story thread

  1. 2026-10-29 Mistral AI releases open weights for Mistral Large 4 ('Le Chonk'), a 1T-parameter multimodal MoE with 49B active per token, trained in its own European data center
  2. 2026-10-29 Mistral Large 4 follow-up: an improved release with more RL after preview, pitched for cyber defense, scores 48% on Coding Agent Index (reported)

More about Mistral AI