Mistral Large 3 675B Instruct 2512 is a multimodal mixture-of-experts language model with 41 billion active parameters drawn from a 675 billion total parameter pool, paired with a 2.5 billion parameter vision encoder that allows the same checkpoint to read images and text together. The granular MoE design is described as state-of-the-art general-purpose on the model card, and the instruct post-training in FP8 targets chat, agentic, and instruction-following use cases rather than raw pretraining. The weights are released openly, with separate FP8, NVFP4, and BF16 builds so that the same model can run on a single node of B200 or H200 accelerators in FP8, on H100 or A100 clusters in NVFP4, or in full precision when memory allows.

The model is positioned for production-grade assistants, retrieval-augmented systems, scientific workloads, and complex enterprise pipelines that benefit from a long context window and native tool use. Multilingual coverage spans English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, and Arabic, and the release notes highlight strong system-prompt adherence along with best-in-class agentic capabilities delivered through native function calling and structured JSON output. In independent head-to-head scoring on shared indexes, the Mistral Large 3 675B variant leads its LLM Stats Score and reasoning index by a wide margin over much smaller vision-language competitors, reinforcing its fit for reasoning-heavy and tool-driven applications rather than lightweight chat.