Hugging Face publishes Mistral Large 4 open weights

On October 29, Hugging Face published official FP8 and NVFP4 weights for Mistral Large 4, giving developers a downloadable route to self-host or adapt the model after its hosted public preview. The NVFP4 checkpoint targets NVIDIA Blackwell GPUs.

Mistral describes Large 4 as a mixture-of-experts model with about 1 trillion total parameters and 49 billion active parameters per token. The model accepts text and images, produces text, and supports a 1 million-token context window.

According to Mistral, the NVFP4 checkpoint combines NVFP4 mixture-of-experts weights with FP8 linear layers. The company says it uses substantially less memory and runs faster than the FP8 version while delivering similar evaluation results. Hugging Face also lists community GGUF conversions from Unsloth and the llama.cpp team, separate from Mistral’s two official checkpoints.

The weights are available under the Mistral Ecosystem Open-Weight License, a modified MIT license that permits use, copying, modification, distribution, further training and derivative works, subject to its terms and notice requirements. A licensee or affiliate operating a Model-as-a-Service business must obtain a separate Mistral commercial license before using the model or a derivative commercially. That requirement does not apply to internal use or use through Mistral’s official products. The license also excludes embedded end-user features and simple request relaying from its definition of Model as a Service.

Hugging Face announced day-one inference support through Transformers, vLLM and llama.cpp, while Mistral’s model cards also document SGLang. The model offers low, medium and high reasoning-effort settings, with high as the default. Hugging Face says reasoning cannot be disabled.

Mistral reports a 48% score on the Artificial Analysis Coding Agent Index and says Large 4 used 47% fewer tokens per task than GLM-5.3. Mistral and Hugging Face also cite scores of 82% on a vulnerability reproduction and patching test, 93% on a 40-task Cybench suite and 93.3% attack resistance on Lakera’s B3 benchmark. The model’s issuers presented those performance and security results; they were not independently reproduced for this story.