Mistral — локальный запуск LLM на своём железе

★ 9.5k

mistral-src is Mistral AI's reference repository with minimal inference code for running the open-weight Mistral family locally, without heavy frameworks. Reach for it when you need to stand up a Mistral model on your own hardware, understand how inference and sliding-window attention work from clean, readable code, or use it as a base for your own integration. It covers the lineup from Mistral 7B and Mixtral 8x7B/8x22B to Codestral, Mathstral, Mistral Nemo, and the multimodal Pixtral models; weights download directly, and you run it from the command line or Python. It is a compact reference implementation for running and studying the models, not a high-throughput production server: for speed and batching in production people use vLLM, and for a simple local chat with no code, Ollama. The officially installable package lives in mistral-inference.