Mistral AI, Europe’s leading artificial intelligence startup, has announced the release of its most ambitious product suite yet: the Mistral 3 family, which includes 10 open-source models that can run on any device from smartphones to autonomous drones and cloud systems.
Mistral 3 family of artificial intelligence models incl Mistral Large 3 and a set of smaller models called Ministry 3 are optimized for edge computing applications. All of these models are released under the Apache 2.0 license, which represents a different and more open-source approach than the open source models offered by Google, OpenAI, and Anthropic.
Introduction of Mistral 3 open source artificial intelligence models
The flagship model of this startup, named Mistral Large 3, is a “combination of experts” (MoE) architecture with 41 billion active parameters from the pool. 675 billion parameters This model is removed, used. This model can text and image Slow processing and its context window up 256 thousand tokens is It is also said that this model is specially designed with Non-English languages Trained, which is a rare occurrence among pioneering models.
According to benchmarks published by Mistral, the Mistral Large 3 (675B) outperformed the DeepSeek-3.1 (670B) and Kimi-K2 (1.2T) in the MMMLU and GPQA-Diamond tests. Also, in AMC and SimpleQA tests, it appeared better than the Dipsik model, but it was slightly weaker than Kimi-K2.

The smallest models called Ministral 3 on devices with at least 4 GB of video memory They are executable, so they can be run on laptops, smartphones and embedded systems and do not require internet. 9 Ministral 3 models in three sizes (3, 8 and 14 billion parameters) and three different types are made for different functions. These three types are:
- Base models for personalization
- Adjusted models with commands for normal chat and task completion
- Optimal reasoning models for complex logical thinking and step-by-step execution of operations
The new Mistral 3 models are available from today in various services such as Mistral AI Studio, Amazon Bedrock, Azure Foundry and Hugging Face (Large 3 and Ministral).
RCO NEWS



