Mistral AI releases the Mistral 3 family, including Mistral Large 3

Background

Mistral AI is a company that develops large language models and offers them through its own products and an API. By late 2025 the company had released several model lines, and the Mistral 3 family was its next generation. Mistral announced it as a single package: one large flagship model and a set of smaller models.

What happened

On 2 December 2025 Mistral AI released Mistral 3, which consists of Mistral Large 3 and three Ministral 3 models at 3B, 8B and 14B parameters ^[1]. The company's announcement is the single source for the details in this entry ^[1].

Mistral Large 3 is a mixture-of-experts model with 675B total parameters and 41B active parameters ^[1]. In this design the model contains many expert sub-networks, and only a few are used for each token. The total is the size of the whole model; the active number is what runs for each word or token processed. This lets a very large model use less computation per token than a dense model of the same total size.

All of the models were released under the Apache 2.0 licence ^[1]. The licence is permissive, allowing use, modification and redistribution within its terms. The release therefore put a flagship model and a family of smaller models under the same open terms.

Key details

Large 3 was trained on 3,000 NVIDIA H200 GPUs ^[1]. Training a model of this size needs a very large cluster of chips working in parallel.

On the LMArena leaderboard, a public ranking based on people comparing answers from different models, Large 3 debuted at #2 among open-source non-reasoning models ^[1]. A non-reasoning model answers directly, without a separate step of working through its reasoning before replying.

The smaller Ministral 3 family includes a reasoning model at the 14B size, which scored 85% on AIME 2025 ^[1]. AIME is a set of competition mathematics problems. This score is one measure of one kind of task and does not summarise the models' overall quality.

Impact

The release gave developers an open-weight model at the top of Mistral's lineup, along with smaller models that could run on less powerful hardware. Because the licence is Apache 2.0 for all the models, a team could choose the size that fitted its budget and use the same terms across the family. For readers following the field, the release is an example of a company publishing a flagship model with open terms. Large 3 is also the first Mistral Large model with fully open weights.