Mistral has published the weights of Mistral Large 4, the model it calls Le Chonk, after a public preview earlier in October. In its announcement, the company says the retrained model is the best open-weight model for coding and among the most capable open models overall, with a 1M context length and three selectable reasoning modes.
It is a 1-trillion-parameter mixture of experts that activates 49 billion parameters per token
Mistral’s model card puts the active count at 49 billion parameters per token, or 52 billion with the embedding and output layers, and describes 128 routed experts with four active per token plus one shared expert, text and image input with text output, and a 1M context length. Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. A second, NVFP4 checkpoint targets Blackwell GPUs with a smaller memory footprint, and Mistral recommends vLLM for serving.
The license is the Mistral Ecosystem Open-Weight License, a modified MIT license that allows commercial and non-commercial use. Operating a model-as-a-service business requires a separate commercial license from Mistral.
Mistral’s own figures show a lead on one private coding benchmark and a tie with GLM-5.3 on another index
Mistral says Large 4 scores 19.7% on Cognition’s private FrontierCode 2.0 benchmark, the lead among the open models it compares. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3, while using 47% fewer tokens per task. On Terminal-Bench 4.0 it reaches 41%, 13 points above the preview, with 30% fewer output tokens.
Running it yourself means storing roughly a terabyte of weights, however few parameters fire per token
Only 49 billion parameters are active per token, but the checkpoint holds all of the roughly 1 trillion. The main checkpoint is split into 54 files of about 20GB each, which comes to roughly 1TB before any quantization. That is well beyond any single consumer GPU, and Mistral’s own vLLM example spreads the model across eight GPUs with a tensor-parallel size of 8.
The NVFP4 checkpoint has a smaller footprint, but Mistral gives no file size or minimum hardware for it.