Un Ministral, des Ministraux | Mistral AI

This post introduces two new models, Ministral 3B and Ministral 8B, designed for on-device and edge computing. It details their performance improvements over existing models in the sub-10B category, highlighting their capabilities in knowledge, commonsense, reasoning, and function-calling. The post also discusses their efficiency, context length support (up to 128k, with 32k on vLLM), and specific architectural features like interleaved sliding-window attention in Ministral 8B for faster inference. It outlines use cases such as on-device translation, smart assistants, local analytics, and robotics, as well as their role in agentic workflows. Benchmark tables and figures are provided to demonstrate performance against competitors. Availability and pricing for API access are detailed, along with licensing information for self-deployment and research use.