Today, Mistral launched a new open-source family of Mistral 3 models. The release includes 10 models, including a large flagship model with multimodal and multilingual support, as well as nine smaller models.
Today, Mistral launched a new open-source family of Mistral 3 models. The release includes 10 models, including a large flagship model with multimodal and multilingual support, as well as nine smaller models.
The flagship Mistral model, dubbed Mistral Large 3, reaches the level of important functionality offered by larger closed-source AI models, including OpenAI's GPT-4o and Google's Gemini 2, and also successfully competes with several open-source competitors, TechCrunch writes .
Large 3 is one of the first open leading-edge models to combine multimodal and multilingual capabilities in a single solution, putting it on par with Meta's Llama 3 and Alibaba's Qwen3-Omni.
Many other companies are now complementing their impressive large language models with separate smaller multimodal models. Mistral has previously used this approach in the Pixtral and Mistral Small 3.1 models.
Large 3 also features a “granular Mixture of Experts” architecture with 41 billion active parameters and 675 billion total parameters, enabling efficient processing in a context window of 256,000 tokens.
This architecture provides both high speed and productivity, making it possible to process voluminous documents and act as an agent assistant for complex enterprise tasks.
Mistral sees Large 3 as suitable for documentation analysis, programming, content creation, AI assistants, and workflow automation.
The line includes 9 different high-performance dense models, presented in three sizes (14 billion, 8 billion and 3 billion parameters) and three variants: Base (pre-trained base model), Instruct (optimized for chat, conversations and assistant-style workflows) and Reasoning (optimized for complex logic and analytical tasks).
Mistral notes that this range of models gives developers and businesses the flexibility to select models according to their specific performance requirements, whether it's maximum speed, cost-effectiveness, or specialized capabilities.
The company claims that the Mistral 3 performs at or above other leading open-source models, while providing greater efficiency and generating fewer tokens for similar tasks. All versions support image processing functions, work with 128K–256K context windows, and operate in multiple languages.
Mistral 3 is capable of running on a single GPU, enabling it to be deployed on cost-effective hardware, from local servers to laptops, robots, and other peripherals with potentially limited connectivity. This is a significant consideration not only for enterprises that store data on-premises, but also for students who need offline access to information, or robotics teams working in remote environments.
