Mistral Large 4 ships as open weights on Hugging Face

Mistral has released Mistral Large 4, nicknamed Le Chonk, as open weights. Hugging Face announced the release in a post dated October 29, 2026.

What the model is

Mistral Large 4 is a multimodal, multilingual Mixture-of-Experts model. It accepts text and images and produces text. The post gives about 1T total parameters, 49B of them active per token, and a 1M-token context window. Official checkpoints come in two variants, FP8 and NVFP4. Unsloth and the llama.cpp team have published GGUF versions.

The license is the Mistral Ecosystem Open-Weight License, described as an MIT derivative that limits model-as-a-service offerings. Anyone planning to build a paid product on top of the weights should read it first.

Capabilities reported

The post positions the model for cybersecurity, coding agents and multimodal work. The figures below are those quoted in the post, and several come from Mistral's own reporting:

• Rank: fifth-best open-weight model on the Artificial Analysis Intelligence index as of October 29, 2026.
• Coding: 61.7% on DeepSWE v1.1 and a 49.8% Coding Agent Index score.
• Business workflows: 59.9% on AutomationBench, which covers business workflows.
• Security: 82% on the Artificial Analysis Cyber Index task of reproducing and then patching a real vulnerability, and 93.3% attack resistance on Lakera's B3 prompt-injection benchmark.

The post says some closed models score near zero on the cyber task because they refuse it. It argues that open weights leave the usage policy and the deployment with the user. The model still declines harmful cyber requests more often than other open-weight models, according to the post.

Running it

Day-0 support covers transformers, vLLM and llama.cpp, plus fine-tuning through TRL. Some practical limits from the post:

• The NVFP4 checkpoint in vLLM needs NVIDIA Blackwell GPUs.
• The Hugging Face Inference Endpoints recipe uses eight H200 GPUs with FP8 weights, serves up to 131K tokens of context and handles 16 concurrent requests.
• On that setup a single request generates about 65 tokens per second. Across 1,000 ShareGPT conversations the total was about 480 tokens per second.
• The 2-bit llama.cpp quants fit on one Apple Silicon Ultra machine with 512 GB of unified memory. Hugging Face measured about 15 tokens per second of generation on an M3 Ultra.
• Reasoning effort can be set to low, medium or high. The default is high and it cannot be disabled.

The post also lists a Dell Enterprise Hub recipe for Dell servers with NVIDIA H200, B200, GB200 or GB300 GPUs, and serverless access through Inference Providers.

What is not in the post

The post gives no pricing for hosted use. The benchmark results are not independently verified in the text, apart from the Artificial Analysis index rankings it cites.