Mistral has released the weights of Mistral Large 4, the trillion-parameter multimodal model it previewed earlier this month under the nickname “Le Chonk”. The French AI lab says the open release follows three extra weeks of reinforcement learning that pushed the model to the front of open-weight coding and defensive cybersecurity. The company published the news on its blog on October 29, 2026, and Hugging Face confirmed the checkpoints are live the same day.
The move matters because a model of this size — about 1T total parameters — is rarely handed out in full. Mistral trained it from scratch on European infrastructure, and says the release gives companies and public institutions a foundation they can own, customize and host themselves instead of renting access from a closed provider.
Mistral Large 4 specs and architecture
Mistral Large 4 is a multilingual Mixture-of-Experts model that combines a large pool of experts with sparse attention. Its text decoder has 72 layers, a hidden dimension of 6,144 and 64 attention heads. Each token activates four of 128 experts plus one shared expert, and the attention indexer selects up to 2,048 key positions per query.
According to Hugging Face, the model ships in two official variants, with GGUFs built by Unsloth and the llama.cpp team. Support is listed for Transformers, vLLM and llama.cpp, with fine-tuning through TRL and deployment options including Inference Endpoints and the Dell Enterprise Hub.
Mistral says the model was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, in the company’s own data centers in Europe.
- 1T total parameters, 49B active per token, Mixture-of-Experts (MoE) design
- Native multimodal: text and images in, text out, via a 1.6B Pixtral vision encoder
