Mistral Large 3 was announced on December 2, 2025 as part of the Mistral 3 generation and is positioned as Mistral's most capable model to date. It is a sparse mixture-of-experts architecture trained with 41 billion active parameters and 675 billion total parameters, marking Mistral's return to MoE design after the Mixtral series. The model is released under the Apache 2.0 license, placing it among frontier instruction-fine-tuned open-weight options. Mistral describes it as achieving parity with the best instruction-tuned open-weight models on general prompts while adding image understanding capability.

For practical deployment, Mistral Large 3 is available across multiple routing options including Cortecs, Mistral AI direct, and OpenRouter, allowing teams to compare latency and throughput on the same underlying weights. Independent inference benchmarks have flagged a recent time-to-first-token increase on the Cortecs-hosted endpoint, suggesting developers with latency-sensitive workloads should verify routing before committing to that path. The combination of a sparse MoE design, open weights, and broad multimodal support makes the model a flexible choice for teams that want frontier-class instruction following with the freedom to self-host or select alternative inference providers.