Mistral unveils Mistral 3, the latest generation of open AI models. What can they do?

The French company Mistral AI has unveiled the Mistral 3 line, an updated generation of open, multimodal, and multilingual models. The line includes three “lightweight” Mistral models (3B, 8B, and 14B) and the flagship Mistral Large 3. This news was announced on the company’s official website.

To make the models as accessible as possible, Mistral collaborated with NVIDIA, Red Hat, and vLLM. 

What can Mistral Large 3 do?

Mistral Large 3 is a model with a mixed expert architecture that was trained on NVIDIA H200 chips. 

It has 41 billion active parameters and 675 billion total parameters. It demonstrates a high level of understanding of text and images and executes instructions in English and over 40 other languages with good quality.

Other capabilities of the model include:

• Answering questions about the content of a specific document.
• Extracting and summarizing information.
• Document analysis and drawing analytical conclusions.
• Querying and comparing multiple documents.
• It also provides context-aware responses that take the entire document into account.

What can the other Ministral 3 models do?

The Ministral 3 series consists of three models (3B, 8B, and 14B), each of which has three versions: Base, Instruct, and Reasoning.

The company claims these models offer the best price-to-quality ratio on the market because they generate significantly fewer tokens with the same level of accuracy.

The three co-founders of Mistral AI, Arthur Mensch, Timothy Lacroix, and Guillaume Lampl, became the first artificial intelligence billionaires in France after receiving funding in September 2025, valuing the company at €11.7 billion ($13.7 billion).