Mistral Large 3
A 675B-parameter open-weight mixture-of-experts multimodal model from Mistral AI
Mistral Large 3 is a general-purpose multimodal mixture-of-experts model with 675 billion total parameters and 41 billion active per token, built by Mistral AI and trained from scratch on 3,000 NVIDIA H200 GPUs. Released under Apache 2.0 in base and instruction-tuned forms, it handles text, images and dozens of languages, and is aimed at assistants, retrieval systems and enterprise workflows.
- 675B total parameters, 41B active per token
- Trained from scratch on 3,000 NVIDIA H200 GPUs
- Apache 2.0 licence on base and instruct checkpoints
- 256k context window with image understanding
- Debuted at #2 among open-source non-reasoning models on LMArena
What it is
Mistral Large 3 is the largest model in Mistral AI's Mistral 3 generation. It is a sparse mixture-of-experts design with 675 billion parameters in total, of which 41 billion are active for each token generated. The company describes it as a general-purpose multimodal model: it reads text and images and produces text, with native function calling and JSON output for tool use. Both a base version and an instruction fine-tuned version are published under the Apache 2.0 licence. Mistral AI positions it for chat assistants, retrieval-augmented systems, scientific workloads and complex enterprise workflows, and says it is its most capable model to date.
How it came about
Mistral Large 3 arrived as part of the Mistral 3 family, alongside three smaller dense models of 14B, 8B and 3B parameters called Ministral 3. It is the company's first mixture-of-experts model since the Mixtral series, which it calls seminal. Mistral AI says the model was trained from scratch and marks a substantial step forward for its pretraining work. Base and instruction-tuned checkpoints were released together, and the company said a reasoning version would follow. The model was developed with NVIDIA, vLLM and Red Hat on deployment and optimisation.
Inside the model
The core is a granular mixture-of-experts language model: 673 billion parameters in the network, with 39 billion active for each token, plus vision components bringing the total to 675 billion. Training used 3,000 NVIDIA H200 GPUs, and the whole Mistral 3 family was trained on NVIDIA Hopper GPUs with high-bandwidth HBM3e memory. The instruct version is published in FP8, with NVFP4 and BF16 variants also available. A 256k context window is supported. NVIDIA added Blackwell attention and MoE kernels, prefill/decode disaggregated serving and speculative decoding work, and a custom Eagle draft model is provided to speed generation.
What it can do
Mistral AI says that after post-training the model reaches parity with the best instruction-tuned open-weight models on general prompts, with best-in-class multilingual conversation performance outside English and Chinese. It understands images, follows system prompts, and supports function calling and JSON output. The company reports it debuted at #2 among open-source non-reasoning models and #6 among open-source models overall on the LMArena leaderboard. Documented limits: it is not a dedicated reasoning model and can be beaten on strict reasoning tasks, it trails vision-first models on multimodal work, and its size makes efficient deployment hard under tight resources.
How to use it
Weights are downloadable from Hugging Face under Apache 2.0, in FP8, NVFP4 and BF16 formats, for both commercial and non-commercial use. The FP8 checkpoint runs on a single node of B200s or H200s; NVFP4 fits a single node of H100s or A100s. Serving is recommended through vLLM 1.12.0 or later. The model is also offered through Mistral AI Studio and platforms including Amazon Bedrock, Azure Foundry, IBM watsonx, Modal, OpenRouter, Fireworks, Unsloth AI and Together AI, with NVIDIA NIM and AWS SageMaker listed as coming soon.
Reception and debate
Mistral AI reports that Mistral Large 3 entered the LMArena leaderboard at #2 in the open-source non-reasoning category and #6 among open-source models overall. The company presents the release as bringing closed-source-level results with the transparency of open weights. No independent reviews, disputes or incidents are described here.
Why it matters
Mistral Large 3 is Mistral AI's first sparse mixture-of-experts model since the Mixtral series and, in the company's words, a substantial advance in its pretraining. Releasing base and instruct checkpoints under Apache 2.0 is intended as a foundation for enterprise and developer customisation, and Mistral AI offers custom training services on top. Joint work with NVIDIA, vLLM and Red Hat produced optimised checkpoints and inference support across frameworks, and the company frames the wider Mistral 3 family as spanning 3B to 675B parameters from edge devices to enterprise deployment.
Questions people ask about Mistral Large 3
What kind of model is Mistral Large 3?
It is a general-purpose multimodal mixture-of-experts model from Mistral AI. It reads text and images and outputs text, with native function calling and JSON output for tool use. The company positions it for chat assistants, retrieval systems, scientific workloads and enterprise workflows.
How large is it and how much runs at once?
The network holds 675 billion parameters in total, but only 41 billion are active for each token generated. The core language model accounts for 673 billion parameters with 39 billion active, with the rest covering vision and related components.
Can I download and use it commercially?
Yes. Both base and instruction-tuned checkpoints are released under the Apache 2.0 licence, which permits commercial and non-commercial use and modification. FP8, NVFP4 and BF16 formats are published on Hugging Face, and the model is also served through several cloud platforms.
What hardware do I need to run it?
The FP8 checkpoint fits a single node of B200s or H200s, while the NVFP4 version runs on one node of H100s or A100s. Mistral AI recommends serving it with vLLM 1.12.0 or later, and a custom Eagle draft model is available for speculative decoding.
What are its main weaknesses?
The published model card notes three limits: it is not a dedicated reasoning model and can lose to such models on strict reasoning tasks, it can trail models built specifically for vision on multimodal work, and its size and architecture make efficient deployment difficult with constrained resources.
Good reading and sources
There is no English Wikipedia article on Mistral Large 3 yet, so the facts on this page come from the sources listed above: mistral.ai/news/mistral-3, huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.