The French AI startup Mistral has developed its first multimodal AI model. Pixtral 12B is based on NeMo 12B and includes twelve billion parameters.

Mistral AI is launching its first AI model that can process both images and text. Pixtral 12B is Mistral AI’s new and first multimodal model, based on one of Mistral AI’s text models, NeMo 12B. The company only announced a link via Hugging Face where developers can download Pixtral 12B. Nothing has been communicated about the technical details and exact licensing conditions yet.

Multimodal model

In July, the company announced its largest AI model to date, Large 2, which will compete with other existing LLMs. What this model lacks is the multimodal aspect. This is what Mistral AI is now trying to offer with the new Pixtral 12B. This model is based on Mistral’s NeMo 12B, is about 24 GB in size and contains twelve billion parameters.

Pixtral 12B can answer questions about any number of images of any size, as well as specific image URLs or images encoded with Base64, the binary-to-text encoding scheme. It is not yet clear what image data Mistral will use to develop Pixtral 12B.

Availability

The model is available on Hugging Face, where it can be downloaded, refined and used under Mistral’s standard license. Reportedly, commercial applications would require a paid license unless they are for research and academic purposes. The company has not yet communicated the technical details and exact licensing terms.

Mistral AI hasn’t been around for very long and has already developed a number of AI models that are helping them secure their place in the AI ​​landscape. For example, the AI ​​company launched NeMo in collaboration with Nvidia, and not much later, the Large 2 AI model saw the light of day. In addition, Mistral is also targeting developers with the launch of Codestral, the AI ​​model for coding.