Mistral AI rolls out full suite of Apache-licensed models

"Mistral AI has released a suite of open source models under the Mistral 3 banner, aiming to scale from a mobile device or drone up to multi-GPU datacenter beasts. While the French company does not share its training data, the decision to open source the models under the Apache 2.0 license is notable. "Open sourcing our models is about empowering the developer community and really putting AI in people's hands, allowing them to own their AI future," Mistral said."

"Mistral Large 3 is the big brother of the lineup and has been trained on a variety of languages, meaning non-English speakers can employ it. "Most AI labs focus on their native language, but Mistral Large 3 was trained on a wide variety of languages, making advanced AI useful for billions who speak different native languages," the firm said. While other AI platforms also claim multilingual abilities, most tend to be optimized for English (and only "likely" to reply in the language of the prompt), as with OpenAI's models."

"Mistral AI boasted of the scalability (only the most relevant experts activate per task), efficiency (processing is distributed over specialized sub-models), and adaptability of its mixture of experts (MoE) architecture, but Mistral 3 is its most flexible development. Mistral 3 has models small enough to run on drones, mobile devices, or laptops. According to the company, there are nine models across three sizes (14B, 8B, and 3B parameters) and three variants: a pre-training Base, a chat-optimized Instruct, and Reasoning with complex logic."

Mistral 3 is a family of open-source models released under the Apache 2.0 license that span from small, edge-capable models to large MoE systems for multi-GPU datacenters. The lineup includes nine models in three sizes (14B, 8B, 3B) and three variants: Base, Instruct, and Reasoning. Mistral emphasizes multilingual training, claiming broader language coverage for non-English speakers. The mixture-of-experts architecture activates only relevant experts per task to improve scalability and efficiency. Smaller models can run on single GPUs, mobile devices, drones, or laptops, reducing hardware costs and enabling offline or edge deployment.

Read at Theregister