Mistral Large 3 is the flagship release within the Mistral 3 family, designed as the provider's most capable open-weight model to date. It is built as a sparse mixture-of-experts architecture with 41 billion active parameters drawn from a much larger pool of 675 billion total parameters, giving it the efficiency profile of a smaller active model while retaining the representational capacity of a very large one. The model was trained from scratch on roughly 3,000 NVIDIA H200 GPUs, a scale that signals a substantial pretraining investment and positions it as a notable step forward in the provider's open-source lineage following the earlier Mixtral series. Mistral AI released Mistral Large 3 under the Apache 2.0 license alongside three smaller dense Ministral variants at 14B, 8B, and 3B parameters, framing the broader family as a performance-to-cost leader across deployment sizes.

After post-training, Mistral Large 3 reaches parity with the strongest instruction-tuned open-weight models on general prompts, while also demonstrating image understanding and best-in-class performance on coding tasks. That dual capability makes it well suited to production assistant applications that need both reasoning over text and grounded multimodal inputs, as well as developer workflows where strong code generation and tool interaction matter. Its combination of a very long context window, open-weight availability, and frontier-tier benchmark standing gives teams a practical option when they want SOTA-adjacent quality without closed-model constraints, and it integrates naturally into pipelines that need fine-grained control over deployment, prompting, and structured outputs.