Announcing Pixtral 12B

Mistral AI · 1 year ago

Mistral released Pixtral 12B, a multimodal AI model combining a 400M parameter vision encoder with a 12B parameter text decoder that processes both images and text. The model achieves 52.5% on the MMMU reasoning benchmark and shows a 20% relative improvement over comparable open-source models in instruction-following tasks. Pixtral is available under Apache 2.0 license via La Plateforme API, Le Chat interface, and local inference through mistral-inference.