The French company Mistral AI has released a new series of open artificial intelligence models Mistral 3, which support multimodality and multilingualism. The flagship model, Mistral Large 3, offers enhanced capabilities in text and image recognition and processing, and cooperation with tech giants is underway for maximum access to innovations.
Mistral Large 3 is the flagship product with 41 billion active parameters. The model was trained on NVIDIA H200 chips and can understand and execute instructions in more than 40 languages. Key features of the model include answering questions, searching and summarizing information, document analysis, and context-dependent responses. To reduce computational costs, Mistral AI collaborates with NVIDIA, Red Hat, and vLLM to optimize performance and pricing.
The Ministral 3 series includes three models (3B, 8B, 14B), each with variations for basic instruction and analytical thinking. These models are considered highly efficient due to the reduction in token generation without quality loss, contributing to their wide application in various fields.
The co-founders of the startup, Arthur Mensch, Timothé Lacroix, and Guillaume Lample, became the first French billionaires in the field of artificial intelligence. Their success is partly explained by recent funding, which raised the company’s value to €11.7 billion.
| Model | Parameters | Main Capabilities |
|---|---|---|
| Mistral Large 3 | 41 billion | Text and visual recognition, multilingual processing, document analysis |
| Ministral 3B | basic configuration | Basic instructions, analytics |
| Ministral 8B | medium configuration | Enhanced text understanding and analytics |
| Ministral 14B | high configuration | Advanced instructions, intensive analysis |
