Mistral releases Pixtral 12B, its first multimodal model

Mistral logo on computer

Image Credits: Rafael Henrique/SOPA Images/LightRocket

French AI startup Mistral has released its first model that can process images as well as text.

Called Pixtral 12B, the 12-billion-parameter model is roughly 24GB in size. Parameters roughly correspond to a model’s problem-solving skills, and models with more parameters generally perform better than those with fewer parameters.

Built on one of Mistral’s text models, Nemo 12B, the new model can answer questions about an arbitrary number of images of an arbitrary size given either image URLs or images encoded using base64, the binary-to-text encoding scheme. Similar to other multimodal models such as Anthropic’s Claude family and OpenAI’s GPT-4o, Pixtral 12B should — at least in theory — be able to perform tasks like captioning images and counting the number of objects in a photo.

Available via a torrent link on GitHub and Hugging Face, the AI and machine learning development platform, Pixtral 12B can be downloaded, fine-tuned and used presumably under Mistral’s standard dev license, which requires a paid license for any commercial applications, but not for research and academic uses.

Mistral hasn’t clarified yet exactly which license applies to Pixtral 12B, however. The startup offers some — but not all — models under an Apache 2.0 license without restrictions. We’ve reached out to Mistral’s PR for more information and will update this post if we hear back.

This writer wasn’t able to take Pixtral 12B for a spin, unfortunately — there weren’t any working web demos at the time of publication. In a post on X, Sophia Yang, head of Mistral developer relations, said Pixtral 12B will be available for testing on Mistral’s chatbot and API-serving platforms, Le Chat and Le Platforme, soon.

It’s unclear which image data Mistral might have used to develop Pixtral 12B.

Most generative AI models, including Mistral’s other models, are trained on vast quantities of public data from around the web, which is often copyrighted. Some model vendors argue that “fair use” rights entitle them to scrape any public data, but many copyright holders disagree, and have filed lawsuits against larger vendors like OpenAI and Midjourney to put a stop to the practice.

Pixtral 12B comes in the wake of Mistral closing a $645 million funding round led by General Catalyst that valued the company at $6 billion. Just over a year old, Mistral — minority-owned by Microsoft — is seen by many in the AI community as Europe’s answer to OpenAI. The younger company’s strategy thus far has involved releasing free “open” models, charging for managed versions of those models, and providing consulting services to corporate customers.

发表回复

您的电子邮箱地址不会被公开。 必填项已用 * 标注