Mistral AI launches Mistral 3 with open models

Mistral AI introduces Mistral 3, a family of open artificial intelligence models that understands text and images and works in more than 40 languages. The lineup includes small models for computers and local devices, as well as a large-scale model designed for business tasks and advanced applications.

All are released under the Apache 2.0 license, which allows you to use, modify and adapt them with few restrictions. For developers and companies, that means more control over their systems and less dependence on a single closed platform.

One large model and three options for modest hardware

The centerpiece is Mistral Large 3, a mixture of experts model. This architecture does not activate all its parameters for every task, only the most relevant ones, which can reduce operating costs.

The model has 675 billion parameters in total, although it activates around 41 billion for each response. Mistral says it trained the model from scratch using 3,000 NVIDIA H200 GPUs and that, after instruction tuning, it reaches a level similar to the best open models available for general queries.

Mistral Large 3 can also interpret images and is designed for multilingual conversations, especially in languages other than English and Chinese. According to Mistral, it debuted in second place in LMArena's category for open models without reasoning and in sixth place in its overall ranking of open models.

The company is releasing both the base version and the instruction-tuned version. A reasoning-focused variant will arrive later.

Alongside Large 3 comes the Ministral 3 series, designed to run AI on smaller servers, personal computers, phones, robots and other devices. It includes models with 3 billion, 8 billion and 14 billion parameters.

Each size comes in base, instruction and reasoning versions. All can work with text and images and are offered under Apache 2.0.

Lower costs for real-world tasks

Model size is not the only factor that makes an application more expensive. It also matters how many tokens, the units AI uses to process text, it needs to generate to complete a task.

Mistral says its Ministral 3 models match or outperform comparable models and that, in some cases, they produce up to ten times fewer tokens. In practice, that can mean cheaper and faster responses for functions such as:

• Summarizing documents on a local computer.
• Analyzing images without sending the data to an external service.
• Creating multilingual assistants for a company.
• Running AI functions on robots, laptops or devices with limited resources.

When accuracy is the priority, reasoning variants can spend more time solving a problem before responding. Mistral says its 14 billion parameter model scored 85% on AIME 2025, a mathematics test, although that result comes from a specific evaluation and does not guarantee the same performance on every task.

Available on several platforms

Mistral 3 is already available through Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI and Together AI. The company also announces upcoming support for NVIDIA NIM and AWS SageMaker.

The infrastructure was developed together with NVIDIA, vLLM and Red Hat. A compressed version in NVFP4 format makes it possible to run Mistral Large 3 on NVIDIA Blackwell NVL72 systems or on a node with eight A100 GPUs or eight H100 GPUs using vLLM.

For you, the most important change is not simply that another large model exists. It is that you can choose between powerful AI for data centers and much smaller versions that run close to where data is generated. That opens up options for controlling costs, protecting sensitive information and adapting the model to specific needs.

Mistral 3 strengthens the company's commitment to open models capable of competing with closed alternatives. What remains to be seen is how much of that performance holds up outside laboratory tests and how easy it is to deploy and maintain these models in real-world applications.