Mistral 3: Four new AI models from France

Mistral 3 is a multimodal AI model family with open weights. It consists of three small (14B, 8B, 3B) and one large model with 675 billion (B) parameters. A special, large reasoning model is expected to follow shortly. As it is a Mixture-of-Experts model, the total of 675 billion parameters are distributed among 41B active parameters each. Expert models are characterized by the fact that only a part of them – the respective expert in the field – answers questions. This makes them faster and more cost-effective. The large model was trained on around 3000 Nvidia H200 GPUs.

All four models are released under the Apache 2.0 license. Mistral writes in the blog post: “Releasing our models in various compressed formats as open source strengthens the developer community and makes AI accessible to everyone through distributed intelligence.” Specifically, this means that with an optimized checkpoint, Mistral Large 3 can be run on Blackwell NVL72 systems and on a single 8×A100 or 8×H100 node with vLLM. The Ministral versions with 3B, 8B, and 14B are said to be optimized for local and edge use, they are multimodal and multilingual. There is a reasoning variant for each. Mistral also speaks of the best price-performance ratio.

Often better than Deepseek in benchmarks

In various benchmarks, Mistral Large 3 performed better than Deepseek V3.1, for example. In the LM Arena, where models compete against each other and are evaluated by humans, the new Mistral model has to concede defeat to the also newly released Deepseek V3.2 – with 1418 points to 1423 points.

Mistral is based in Paris. The company develops its own large language models. The founders of Mistral previously worked at Google's DeepMind and Meta. Mistral is seen as a European hope to compete with Big Tech from the USA. However, there are also other initiatives, such as the Swiss language model Apertus, which was developed by researchers from ETH Zurich, among others.

(emw)