Mistral AI has unveiled its most ambitious model suite yet, with Mistral 3 positioning the French startup to compete directly with OpenAI and Anthropic at the frontier while maintaining its open-weight commitment. The release includes a 675B parameter sparse MoE flagship and nine smaller dense models.
The Mistral 3 Suite
The flagship Mistral Large 3 boasts 675 billion parameters in a mixture-of-experts architecture, described by CEO Arthur Mensch as “fat but sparse.” The company claims it is the world’s best open-weight multimodal and multilingual AI, outperforming rivals across key benchmarks.
Joining the flagship are nine smaller dense models under the Ministral line, ranging from 3B to 14B parameters. All models in the suite are released under the Apache 2.0 license, maintaining Mistral’s tradition of making weights accessible.
Enterprise Focus
The release comes with Mistral Forge, an enterprise platform designed for organizations deploying AI at scale. The company has been building out its commercial offerings while keeping the open-weight commitment that has attracted a strong developer community.
Mistral’s valuation has climbed to approximately $23 billion as the company expands from research into enterprise products. The new suite is designed to serve both the open-source community and enterprise customers seeking alternatives to closed API providers.
Open-Source Strategy
Mistral has consistently pursued a dual strategy: releasing capable open-weight models while building premium API services. Mistral 3 continues this approach, offering organizations the choice between self-hosting the models or using Mistral’s managed API.
The company’s push into the frontier model space comes as competition intensifies among open-weight providers. With Chinese labs like DeepSeek and ByteDance’s Moonshot making significant strides, Mistral’s European heritage positions it as a Western alternative with strong open-source credentials.
The model is available immediately through Mistral’s API and can be downloaded for self-hosting from the company’s platform.