Pixtral
Multimodal AI models you can run locally with Mistral's GPU inference library, which is open source under Apache 2.0 and archived.
Pixtral is a family of Mistral models for developers who want to run multimodal AI on their own hardware. The associated mistral-inference project is archived and no longer maintained. That status applies to the inference library.
The available models include Pixtral 12B, Pixtral 12B Base and Pixtral Large Instruct, with weights distributed through Hugging Face. The base and instruction variants give developers a choice of model within the same family. Mistral's library supports multimodal instruction following and lets applications run models through Python or a command-line interface.
Local inference uses downloaded model weights. The library requires a GPU, so this route is for people with GPU hardware rather than those looking for a CPU-only application. It's a developer library rather than a ready-made chat interface.
Pixtral 12B weights and the inference code use Apache 2.0. Pixtral Large Instruct uses the Mistral AI Research License, which restricts use to the purposes it authorizes; check its terms before commercial deployment. For server deployment, the repository also includes code for a vLLM serving image that uses Transformers in place of Mistral's reference implementation.
