Mistral Publishes Technical Report for Mistral Large 4
Mistral has published a technical report for Mistral Large 4, the 1T-parameter flagship whose weights shipped on October 29. The model activates 52 billion parameters per token using a mixture of experts with 128 experts, four of which are selected per token plus one shared expert. The report details Multi-head Latent Attention, Sparse Attention and KV-cache reuse across layers, the changes that cut memory and compute on long-context inference for a 1M-token window. The model handles images, more than 160 languages, tool calling and agentic workflows, and its weights are on Hugging Face in FP8 and NVFP4.
On the Artificial Analysis Coding Agent Index it scores 48%, about level with GLM-5.3 while using 47% fewer tokens according to Mistral. The model is not runnable on a consumer GPU, so the report is aimed at people working on LLM architecture and optimization rather than at local deployment.
Related: Mistral Publishes Mistral Large 4 Open Weights Under a Modified MIT License, Mistral Unveils Large 4, a 1T-Parameter Open-Weight Model Trained in Europe