Mistral Unveils Pixtral 12B: A Multimodal AI Model

On 11th September, Mistral AI announced its latest advanced AI model capable of processing both images and text. Pixtral 12B, a one of a kind model, employs about 12 billion parameters and is capable of vision encoding, enabling it to interpret images alongside text. It is built on their previous model, Mistral’s Nemo 12B, capable of understanding text, with the addition of features like the 400 million-parameter vision adapter. In a post on X,  Sophia Yang, Head of Developer Relations at MistralAI, shared that the model can be downloaded via a torrent link on GitHub and Hugging Face, and  used under an Apache 2.0 license without restrictions. It will be available on le Chat and la Plateforme soon.

What is Mistral AI?
Based in France, and built by former employees of Meta and Google in 2023, the founders - Arthur Mensch, Guillaume Lample, Timothée Lacroix - aim to make GenAI more fun and accessible. Last year, they closed a seed-stage financing round of over €105 million led by Lightspeed, a US based VC firm. Earlier this year, they held a hackathon in Paris, and provided GPUs for their first hackathon in Europe. Mistral AI follows an open-source approach, releasing all models under open licenses for free use and modification. They focus on creating efficient, accessible models which are trained on diverse datasets—text, code, and images—making them more versatile than those trained on a single type. While still a year old, it competes with the likes of Anthropic PBC’s Claude family, OpenAI’s GPT-4o and Google LLC’s Gemini, among others.