French startup Mistral has launched Pixtral Large, a new monumental model with 124 billion parameters suitable for multilingual OCR, chart understanding and reasoning.
Pixtral Large includes a 1-billion-parameter vision encoder and a 123-billion-parameter decoder, ensuring it can process both text and visual data.
Additionally, the new upgrades will enhance the company's free web-based chatbot, Le Chat, with image generation, web search and an interactive canvas added to its functionalities.
Le Chat will now have the ability to process PDFs, extract data from tables, graphs, and equations, and produce high-quality visuals.
Mistral hopes the features will enable users to use Le Chat as a versatile AI assistant to save time and effort performing tasks that would require multiple tools otherwise, as well as creating a better AI experience to co-design models and product interfaces.
Pixtral Large's weights and model are publicly available but distributed under a Mistral AI Research Licence limited to non-commercial, research-focused applications.
Le Chat can be accessed by users provided they have a Mistral, Google, or Microsoft account.
Mistral AI competes with giants of the industry such as Google and OpenAI, with the latter releasing its own interactive sidebar feature Canvas for ChatGPT, while Mistral has launched Le Chat's Canvas.
A recent survey suggested that usage of Mistral’s models and API by large enterprises remained far behind those of US-based companies such as OpenAI and Microsoft.
However, in the post-presidential election world, there are indications that European options such as Mistral may become more attractive compared to their US counterparts.