Mistral Large 3 is built as a general-purpose multimodal foundation model with a sparse Mixture-of-Experts design, drawing on the lineage of Mistral's earlier Mixtral work while stepping into a much larger scale. With 41B active parameters out of 675B total, it routes each request to a focused subset of experts, aiming to deliver frontier capability without paying the full compute cost at inference time. The model is designed to handle text alongside image and PDF inputs, producing text outputs, and is positioned for broad workloads such as conversational assistants, document analysis, function-calling agents, and structured outputs where both reasoning depth and efficiency matter.
Trained from scratch on roughly 3000 NVIDIA H200 GPUs, Mistral Large 3 represents a meaningful leap in pretraining scale for the family and is released under the Apache 2.0 license, keeping it in the open-weights camp alongside other frontier-grade models. After post-training, it reaches parity with leading instruction-tuned open-weight systems on general prompts while also demonstrating strong image understanding, making it a practical choice for tasks that blend visual and textual reasoning. Its 256k context window supports long-document workflows, and the mixture-of-experts backbone leaves room for further specialization, whether through domain fine-tuning, agent pipelines, or retrieval-augmented setups that benefit from a capable base model with room to grow.
