Pixtral 12B: Mistral's multimodal model with a vision encoder trained from scratch

On September 17, 2024 Mistral AI publishes Pixtral 12B, a model that understands both text and images. Mistral trained its own vision encoder from scratch instead of reusing an existing one. Images are read at their original resolution and aspect ratio, and several images can be mixed with text in the same conversation. The weights are released under the Apache 2.0 license, so they can be used commercially, and the model is also available through the API and in Le Chat.