Mistral AI Unveils 10 New Open-Source Models, Including Multimodal Mistral Large 3 and Edge-Optimized Ministral 3 Series

Paris, France – Mistral AI has officially launched its new "Mistral 3" family of artificial intelligence models, comprising a total of 10 open-source offerings. This release includes the flagship Mistral Large 3, a powerful multimodal and multilingual model, alongside a suite of nine smaller "Ministral 3" models designed for efficient deployment on edge devices. The announcement marks a significant move by the European AI startup to challenge established tech giants with flexible, open-source solutions.

The new lineup is spearheaded by Mistral Large 3, a sparse mixture-of-experts model featuring 41 billion active parameters from a total of 675 billion. This model boasts advanced capabilities in both text and image understanding, supporting a 256,000-token context window. Notably, Mistral Large 3 was trained with a strong emphasis on non-English languages, aiming to make advanced AI accessible to a global audience.

Complementing the large model are the Ministral 3 series, available in 3 billion, 8 billion, and 14 billion parameter variants. These smaller models are optimized for edge computing, enabling deployment on devices ranging from laptops and smartphones to drones and embedded systems, even without cloud connectivity. According to a social media post, the "Ministral 14B looks cracked," suggesting impressive performance for its size, further enhanced by vision capabilities and improved multilingual training data.

All Mistral 3 models are released under the permissive Apache 2.0 license, allowing for unrestricted commercial use and customization. This strategy underscores Mistral AI's commitment to open-source development, positioning itself as an alternative to proprietary systems by prioritizing flexibility, cost-efficiency, and data sovereignty for enterprise customers. The models are immediately available across various platforms, including Mistral AI Studio, Amazon Bedrock, and Hugging Face.