Mistral Releases Mistral Large 4 Open-Weight MoE Model on Hugging Face

Mistral Large 4 (also known as Le Chonk) has been released on Hugging Face as a multimodal, multilingual Mixture-of-Experts (MoE) model. The model accepts text and image inputs to generate text, combining a Pixtral vision encoder with a text decoder utilizing Multi-head Latent Attention and sparse attention mechanisms. The model is released under the Mistral Ecosystem Open-Weight License, establishing official checkpoints in FP8 and NVFP4 alongside community GGUF quantizations. Benchmark results reported by Mistral indicate an 82% score on the Artificial Analysis Cyber Index vulnerability reproduction and patching test, alongside a 49.8% score on the Coding Agent Index. Inference is supported across day-0 frameworks including transformers, vLLM, and llama.cpp, enabling deployment across enterprise server environments and local hardware.

Key takeaways

• Mistral Large 4 is an open-weight Mixture-of-Experts multimodal model that accepts text and image inputs, combining sparse attention with a Pixtral vision encoder.
• Distributed under the Mistral Ecosystem Open-Weight License, official variants include FP8 and NVFP4 alongside GGUF quantizations.
• Reported benchmark scores include 82% on vulnerability reproduction and patching, alongside a 49.8% Coding Agent Index rating.
• Day-0 runtime integrations are available for transformers, vLLM, and llama.cpp, accommodating distributed servers and select local hardware.