Mistral Large 3

What it is

Mistral Large 3 is the flagship model of the Mistral 3 family, released by Mistral AI on 2 December 2025^[1]. It is a sparse mixture-of-experts (MoE) model with 675 billion total parameters and 41 billion active parameters^[1]. It is the company's first mixture-of-experts model since Mixtral^[1].

In a sparse mixture-of-experts model, the network contains many separate "expert" sub-networks, and a routing step sends each token to only a small subset of them. The model can therefore hold far more parameters than the amount of computation spent on any single token — which is why Mistral Large 3's 675 billion total parameters come with an active count of 41 billion.

How it was built

Mistral trained the model from scratch on 3,000 NVIDIA H200 GPUs^[1]. It is not text-only: the model includes image understanding, and it supports more than 40 languages^[1].

Open weights

Both the base and the instruct version were released under the Apache 2.0 license^[1]. Mistral describes this as the first Mistral Large model with fully open weights^[1]. Apache 2.0 is a permissive license that allows use, modification and redistribution, subject to its attribution terms; applying it to a model this size means outside developers can download and run the weights themselves rather than only calling an API. A reasoning version was announced as coming, but was not part of the launch^[1].

Benchmark standing

At its debut, Mistral Large 3 placed second among open-source non-reasoning models and sixth among all open-source models on the LMArena text leaderboard^[1]. LMArena ranks models from blind pairwise comparisons in which users pick the better of two responses; because it is a public, continuously updated ranking rather than a fixed test set, a position on it describes how the model compared with others at that moment.

Running and deploying it

Mistral published an NVFP4 checkpoint that runs on a single node of eight A100 or eight H100 GPUs using vLLM^[1]. NVIDIA added support for the model in TensorRT-LLM and SGLang^[1]. Mistral says it works with vLLM and Red Hat to make Mistral Large 3 accessible to the open-source community^[1].

At launch the model was available on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, IBM watsonx, OpenRouter, Fireworks and Together AI^[1], and it is also available on Modal and Unsloth AI^[1]. Developers reaching it through Mistral's API use the name `mistral-large-2512`^[2], and Mistral's models overview lists it under version 25.12^[2].

Where it sits

Mistral Large 3 was released together with the Ministral 3 models on 2 December 2025^[1], so the family launch covered both the flagship and smaller siblings. The documentation records the API model as deprecated on 31 October 2026, with retirement scheduled for 30 April 2027 and Mistral Large 4 named as its replacement^[2][3].

Taken together, the release marked a shift in how Mistral shipped its largest model: a mixture-of-experts design, multimodal input, a permissive open license and day-one availability across a long list of cloud and inference providers^[1].