Mistral announced Mistral 3 on 2 December 2025 and says the family includes three small dense models at 14B, 8B and 3B parameters plus Mistral Large 3, a sparse mixture-of-experts model with 41B active parameters and 675B total parameters.
Mistral says all Mistral 3 models are released under the Apache 2.0 license, and the Mistral Large 3 release includes both base and instruction-fine-tuned versions.
Mistral says Mistral Large 3 was trained from scratch on 3000 NVIDIA H200 GPUs and that the model supports image understanding and multilingual conversations after post-training.
Mistral says it worked with vLLM and Red Hat on an NVFP4 checkpoint for Mistral Large 3, and the release notes say that checkpoint can run on Blackwell NVL72 systems or on a single 8xA100 or 8xH100 node using vLLM.
Mistral's model overview lists Mistral Large 3 as an Apache 2.0 open-weight, general-purpose multimodal model at version 25.12, and it lists Ministral 3 14B, 8B and 3B as Apache 2.0 models with text and vision capabilities.
Mistral's agents documentation says its Agents and Conversations API supports persistent state across conversations, conversations with base models or one or more agents, and built-in connector tools for code execution, web search, image generation and document libraries.
The same agents documentation says developers can register MCP servers as managed connectors and use them as tools in conversations and agents.
For developer teams, the concrete product move is that Mistral is connecting open-weight model releases, deployment notes and agent-platform APIs in one vendor stack; the cited pages do not provide an independent production cost comparison against other providers.
