Pixtral 12B

Mistral AI has announced the release of Pixtral 12B, its first multimodal model designed to process both images and text. This model features a new 400M parameter vision encoder and a 12B parameter multimodal decoder, allowing it to handle variable image sizes and multiple images within a long context window of 128K tokens. Pixtral 12B excels in multimodal tasks such as document question answering and chart understanding, achieving a score of 52.5% on the MMMU reasoning benchmark, outperforming several larger models. It maintains strong performance on text-only benchmarks, making it a versatile tool for developers. The model is designed to be a drop-in replacement for Mistral Nemo 12B, providing best-in-class multimodal reasoning without sacrificing text capabilities. Pixtral's architecture allows for efficient processing of images at their native resolution, enhancing its ability to understand complex diagrams and documents. The model has been benchmarked against both open and closed models, demonstrating superior performance in instruction following tasks. Pixtral is available for use via La Plateforme and Le Chat, with open-source prompts and evaluation benchmarks to be shared with the community.

- Pixtral 12B is Mistral AI's first multimodal model, integrating image and text processing.

- It achieves a 52.5% score on the MMMU reasoning benchmark, outperforming larger models.

- The model supports variable image sizes and can process multiple images in a 128K token context.

- Pixtral excels in both multimodal and text-only instruction following tasks.

- It is available for use through La Plateforme and Le Chat, with open-source resources planned.