Mistral said on October 29 it is releasing open weights for Mistral Large 4, shifting its flagship multimodal model beyond API-only access. The company says the model remains available through Mistral Studio, its API, and OpenRouter, while teams that want tighter infrastructure control can now download the weights for self-hosting.
The Hugging Face checkpoint is listed as mistralai/Mistral-Large-4-1T-A52B. Mistral describes Large 4 as a roughly 1T-parameter model, and the Hugging Face listing gives it a 1M-token context window. Mistral also says hosted API pricing starts at $0.70 per million input tokens and $2.10 per million output tokens.
For deployment, Mistral points to self-hosted options via Hugging Face, vLLM, and SGLang. Hugging Face says the release includes FP8 and NVFP4 variants, with day-0 support across transformers, vLLM, and llama.cpp.
The licensing boundary is the main catch: Mistral says the weights use its Mistral Ecosystem Open-Weight License, but companies planning to offer Large 4 as a hosted model service need a separate commercial agreement.