French firm Mistral AI has the wind in its sails with the launch of the Mistral 3 model family.
The release includes a series of small models, dubbed “Ministral 3,” and the new flagship model, “Mistral Large 3,” which the company describes as its “most capable model to date.”
Crucially, all models are released under the permissive Apache 2.0 license. And it’s Mistral’s latest attempt to jockey for position in the open-source AI arena.
Frontier open model
The centerpiece of the release is Mistral Large 3, a model that introduces advancements in architecture and capability. This model is Mistral’s first mixture-of-experts (MoE) model since its Mixtral series, representing a leap in pretraining.
The model is built with 675 billion total parameters, but utilizes a sparse MoE architecture, meaning only 41 billion parameters are actively used for any given query. This design delivers scale without commensurate computational cost.
Trained on 3,000 of Nvidia’s H200 GPUs, Mistral Large 3 reportedly “achieves parity with the best instruction-tuned open-weight models on the market on general prompts.”
The model’s capabilities include native multimodal processing, allowing it to understand images, and offer performance on multilingual conversations beyond English and Chinese. It debuts at the #2 position in the open-source non-reasoning models category on the LMArena leaderboard.
Ministral 3
Complementing the large model is the Ministral 3 series, a suite of small, dense models available in three billion, eight billion, and 14 billion parameter sizes. They are designed for edge and local use cases.
Each size is offered in base, instruct, and specialized reasoning variants, all featuring native multimodal and multilingual capabilities. They are designed for environments where computational resources are highly constrained, such as mobile devices, laptops, or edge computing systems.
The company stated that the Ministral instruct variants “match or exceed the performance of comparable models while often producing an order of magnitude fewer tokens.” Furthermore, the reasoning variants, such as the 14B model, have demonstrated accuracy in their weight class, scoring 85% on the AIME ’25 benchmark.
Rise of distributed intelligence
The Ministral family underscores a shift toward “distributed intelligence,” where advanced AI can run locally on consumer devices rather than relying solely on cloud servers.
This has ramifications for data privacy, as sensitive information can be processed on-device without being sent to the cloud. It also ensures resilience in environments with intermittent connectivity, making sophisticated AI available “from data center to robot.”
Industry collaboration
Mistral AI did not achieve this alone. The company announced a collaboration with technology firms Nvidia, vLLM, and Red Hat to ensure the models are immediately accessible across a range of hardware.
The entire Mistral 3 family was trained on Nvidia Hopper GPUs. For inference, the collaboration has resulted in optimized checkpoints, including the NVFP4 format, built with llm-compressor, allowing Mistral Large 3 to run on systems like the Blackwell NVL72 and single 8-GPU nodes.
Nvidia engineers integrated customized kernels and support for frameworks like TensorRT-LLM and SGLang, specifically targeting the complex sparse MoE architecture for “high-throughput” service.
The partnership ensures that the new models are not just research curiosities but are deployable in production from day one, giving developers a path across Nvidia’s entire hardware ecosystem, from data centers to Jetson devices at the edge.
An open future?
By open-sourcing its frontier model family, Mistral AI wants to challenge the established dominance of closed-source systems like those from OpenAI and Google.
The models are available on cloud platforms and repositories, including Mistral AI Studio, Amazon Bedrock, Azure Foundry, and Hugging Face. Furthermore, Mistral AI is offering custom model training services, allowing organizations to fine-tune the models on proprietary datasets for domain-specific and scalable enterprise deployments.
The company’s philosophy is clear: “The future of AI is open.”
The number three is in vogue. ChatGPT recently reached its third anniversary.