Mistral Large 4 ships as open weights: 1T MoE, 49B active, 1M context
Victoria Shinder2 min readThe weights of Mistral Large 4, nicknamed Le Chonk, went up on Hugging Face on 29 October, three weeks after Mistral previewed the model and promised a release by the end of the month. It is a mixture-of-experts model with about one trillion parameters in total and 49 billion active per token, reads text and images, writes text, and has a context window of one million tokens. Unsloth and the llama.cpp team have published GGUF builds alongside the official checkpoints.
The licence is not plain MIT. Mistral calls it the Mistral Ecosystem Open-Weight License, an MIT derivative that sets limits on offering the model as a hosted service.
Architecturally, the Hugging Face write-up describes two separate kinds of sparsity. A router picks four of 128 experts per token plus one shared expert that always runs, and a DeepSeek-style sparse attention indexer chooses up to 2,048 earlier positions for each query on top of Multi-head Latent Attention. The decoder has 72 layers; image input goes through a 1.6-billion-parameter Pixtral encoder.
Mistral leads with security. On Artificial Analysis' Cyber Index, which asks a model to reproduce a real vulnerability in open-source code and then patch it, Large 4 scores 82%; Mistral says several closed models score near zero because they refuse the task. On cyber prompts from JailbreakBench, StrongREJECT and AgentHarm its average refusal rate is reported as the highest among open-weight models. Coding figures include 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, and as of 29 October it ranks fifth among open-weight models on the Artificial Analysis Intelligence Index.
What it means
For teams that cannot send vulnerability data to an API, a reproduce-then-patch model they can host themselves is the point, and open weights let anyone measure the refusal behaviour instead of trusting a description. The benchmark numbers are Mistral's and Hugging Face's selection; the licence terms matter more than the parameter count for anyone planning to resell access.