French AI startup mistral has released its first model that may process each images and text.
The 12 billion parameter model, called Pixtral 12B, is about 24 GB in size. (Parameters roughly correspond to a model's problem-solving capabilities, and models with more parameters generally perform higher than those with fewer parameters.) Available on GitHub and the AI and machine learning development platform Hugging faceThe model could be downloaded, optimized and used under Mistral's standard license. Any industrial application requires a paid license, but no license is required for research and academic purposes.
Pixtral 12B relies on Mistral's text model Nemo 12B and might answer questions on any variety of images of any size, given either image URLs or images encoded with the base64 binary-to-text encoding scheme. Like other multimodal models (e.g. Anthropic's Claude family, GPT-4o, etc.), Pixtral 12B should – at the very least in theory – give you the chance to perform tasks comparable to labeling images and counting the variety of objects in a photograph.
Unfortunately, this creator was unable to check out Pixtral 12B – there have been no working web demos on the time of publication. In a post on X, Sophia Yang, head of development at Mistral, wrote, said that Pixtral 12B can be available for testing “soon” on Mistral’s chatbot and API serving platforms Le Chat and Le Platforme.
It is unclear which image data Mistral might need used to develop Pixtral 12B.
Most generative AI models, including the opposite models of Mistralare trained on massive amounts of public—and infrequently copyrighted—data from across the web. Some model providers argue that fair use entitles them to make use of public data. Many copyright holders disagree and have filed lawsuits against the larger providers, including OpenAI and Midjourney, to place an end to the practice.
The release of Pixtral 12B comes shortly after Mistral closed a $645 million funding round led by General Catalyst, valuing the corporate at $6 billion. Mistral is barely a yr old and is seen by many within the AI community as Europe's answer to OpenAI. The company's strategy to this point has been to release free “open” models, charge for managed versions of those models, and offer consulting services to enterprise clients.