Pixtral 12B

Pixtral 12B is an open-source, natively multimodal vision-language model (VLM) developed by Mistral AI and released on 17 September 2024 under the Apache 2.0 license. It combines a 12-billion-parameter multimodal decoder (based on Mistral NeMo) with a 400-million-parameter vision encoder trained from scratch.

The model was trained on interleaved image–text data, so it understands both natural images and documents: it analyzes charts and figures, answers document questions (DocVQA), performs multimodal reasoning and follows instructions. It supports a 128K-token context window and images of variable size and aspect ratio at native resolution, as well as multiple images at once.

Pixtral 12B is available with open weights on Hugging Face (Apache 2.0), locally via mistral-inference or vLLM, and also in Le Chat and via the API on Mistral’s platform (la Plateforme).

On multimodal benchmarks Pixtral 12B achieves, among others: MMMU 52.5%, MathVista 58.0%, ChartQA 81.8%, DocVQA 90.7% and VQAv2 78.6%. The model has since been marked as deprecated — Mistral has replaced it with newer vision models.