Mistral Large 3
2025-12-02 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI launches Mistral Large 3, its most capable model: a multimodal mixture-of-experts with 675 billion parameters (41 billion active), released under the Apache 2.0 license, with a 256k-token context.
What changes
- First mixture-of-experts model from Mistral since Mixtral, trained from scratch on 3,000 NVIDIA H200 GPUs.
- Announced parity with the best open instruction-tuned models on general prompts, image understanding and the best multilingual performance outside English and Chinese.
- Debuts 2nd among open non-reasoning models (6th among all open models) on the LMArena leaderboard, according to the publisher.
- Base and instruct versions released, NVFP4 checkpoint for an 8×A100 or 8×H100 node with vLLM; a reasoning version is announced.
### Specifications - 675 billion parameters in total, 41 billion active; 2.5-billion-parameter vision encoder; 256k context. - Languages: more than 40 native languages according to the post; text and image inputs. - License: Apache 2.0. - Availability: Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM watsonx, OpenRouter, Fireworks, Together AI, Unsloth; NVIDIA NIM and SageMaker announced. - API identifier: `mistral-large-2512`. ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Mistral Large 3 | 0.5 | 1.5 | ### Benchmarks The charts in the post and on the Hugging Face model card are images; there are no numerical scores in text. The only published reference point: 2nd place among open non-reasoning models (6th among open models) on LMArena at launch. ### Key takeaways An open, permissive flagship model that can be deployed on a single node of recent GPUs, making self-hosting a frontier-class model possible for enterprises.
Links
Tags : LLM, Mistral, Large, Mixture-of-Experts, Open Source, Apache 2.0