Mistral Large 4 (Le Chonk) released on Hugging Face

On October 29, 2026, Mistral Large 4, also known as Le Chonk, was published on Hugging Face. It is a multimodal and multilingual model from Mistral, built on the MoE (Mixture-of-Experts) scheme: only part of the weights is activated for each token. It accepts text and images as input and outputs text. It is claimed to have about 1 trillion total parameters, 49 billion active parameters, and a 1 million token context window. The weights are distributed under the Mistral Ecosystem Open-Weight License — a derivative of MIT with restrictions for those offering the model as a service.

The main difference from previous versions is two types of sparsity at once. First: a large pool of experts, from which a small group is selected for each token. Second: sparse attention with an indexer that decides which earlier token positions to attend to. Added to this is the Pixtral vision encoder. GGUF versions (a quantized weight format for local launch) were prepared by Unsloth and the llama.cpp team — meaning the model can be run not only on server accelerators.

In practice, three strengths are claimed: cybersecurity, coding agents, and multimodality. Inference is supported in Transformers, vLLM, and llama.cpp; fine-tuning is via TRL. Inference Endpoints and Dell Enterprise Hub are mentioned separately. For teams building agents for code work or parsing screenshots and documents, this is a ready-made toolkit for one model. The license is not pure MIT: if you serve the model externally as a service, you should read the restrictions rather than pulling the weights into production out of habit.

What remains unknown: the announcement text has a section about benchmarks, but no specific numbers are given. There is also no pricing for use via Inference Endpoints. The volume of training data and the number of experts are not disclosed — no data.