Annonce de Pixtral 12B

Pixtral 12B is a deprecated multimodal model from Mistral AI, trained natively on interleaved image and text data. It features a 12 billion parameter multimodal decoder and a new 400 million parameter vision encoder that supports variable image sizes and aspect ratios. The model demonstrates strong performance on multimodal reasoning tasks and instruction following while maintaining high performance on text-only benchmarks. With a 128k token context window, Pixtral can process multiple images and was shown to outperform comparable open-source models in various evaluations. The architecture was designed to excel at understanding complex documents and charts without compromising core text capabilities like coding and math.