Mistral Large 3 is presented by Mistral as a state-of-the-art, open-weight, general-purpose multimodal model anchored on a granular Mixture-of-Experts design. The official documentation places 41 billion active parameters within a much larger 675 billion parameter total, a pattern typical of MoE systems where only a slice of the network activates per token. That sparse activation is intended to deliver strong general capability while keeping inference cost more manageable than a comparably sized dense model, and the Apache 2.0 license signals Mistral's intent for the weights to be broadly deployable and self-hostable by organizations that need control over their stack.

General availability was reached on December 2, 2025 under the v25.12 versioning as part of the Mistral 3 generation, and enterprise reach was quickly broadened through IBM watsonx, where the same weights can be consumed as a managed service. The model targets a wide working envelope, with a 256k context window, function calling, structured outputs, and document question answering available through the chat completions endpoint, making it well suited to long-document reasoning, retrieval-augmented workflows, and tool-using agents. The combination of open weights, a long context, and multimodal input gives teams a flexible foundation that can be fine-tuned or deployed as-is depending on their infrastructure and governance needs.