Mistral releases Large 3, an open 675-billion-parameter model
French company Mistral AI has released Large 3, a mixture-of-experts model with 675 billion total parameters, alongside three smaller Ministral 3 models. The entire family is distributed under the Apache 2.0 license.
Mistral AI has unveiled Mistral 3, a new model family that takes its open approach to an unusual scale for Europe. The release includes Mistral Large 3, a multimodal model with 675 billion total parameters, and three compact versions called Ministral 3, with 3 billion, 8 billion and 14 billion parameters.
All of them are released under the Apache 2.0 license, a permissive license that allows the models to be used, modified and redistributed, including in commercial products. For companies and developers, that means they can adapt the weights to their own data and tasks without being required to rely on a closed API.
A large model that activates only part of its capacity
The figure of 675 billion parameters places Mistral Large 3 among the largest open models. But it does not use all of those parameters to answer every request: it is a mixture-of-experts or MoE model, an architecture that divides the model into several specialized blocks and activates only the ones best suited to each piece of text.
In its case, Mistral says that 41 billion parameters are activated per token. The distinction matters because an MoE model can bring together more knowledge and capabilities without the cost of generating each response rising in the same proportion as it would for a dense model, where every parameter is involved in every calculation.
The company says it trained Large 3 from scratch on 3,000 NVIDIA H200 GPUs. It is releasing both the base version—intended for later customization—and the instruction-tuned variant. The model accepts text and images as input and is designed for multilingual conversations, with native support for more than 40 languages.
On the LMArena leaderboard, Large 3 debuts in second place among open models not specifically designed for reasoning and in sixth place among open models overall, according to Mistral. These rankings are a useful reference for comparing responses as perceived by users, but they do not replace testing in a specific use case: programming, internal documents and customer support can produce different results.
From a data center to a local computer
The other half of the announcement is Ministral 3. Mistral offers three sizes—3 billion, 8 billion and 14 billion parameters—and releases base, instruct and reasoning versions for each. The latter are designed to spend more internal steps on complex problems before responding.
The company reports that Ministral 3 14B achieves 85% accuracy on AIME 2025, a competitive mathematics test. That is a striking result for a model of this size, but it is important to distinguish between performing well on a standardized exam and reliably handling complete business processes: data quality, connected tools and human oversight remain decisive.
The smaller models also understand images and are designed to run on local devices or more contained infrastructure. That option matters when an organization wants to analyze documents, images or sensitive information without sending every request to an external service. Mistral highlights optimized deployments for computers with RTX GPUs, DGX Spark systems and Jetson devices, as well as servers.
Openness with a considerable infrastructure bill
An open license does not automatically make Large 3 easy to run. Mistral has published a compressed checkpoint in NVFP4 format and says it can be served on a node with eight A100 or H100 GPUs using vLLM. That is still data center infrastructure, beyond the reach of most individual teams.
That is where Ministral matters: it covers the part of the market where available memory, power consumption and latency really count. The family lets users choose between a lightweight local model and a large-scale system while retaining the same license and similar multimodal capabilities.
Mistral Large 3 is available today on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, IBM watsonx, OpenRouter, Modal, Fireworks, Unsloth AI and Together AI. In Mistral’s API, the announced price is $0.50 per million input tokens and $1.50 per million output tokens. Ministral 3 8B costs $0.15 per million tokens for both input and output.
The announcement puts a European company back in the spotlight in the race for large-scale open models. The next test will be less glamorous than a leaderboard ranking: whether the community can fine-tune, run and integrate Large 3 and the Ministral models easily enough for that openness to translate into real-world applications.
