Mistral Large 3 675B Instruct 2512 NVFP4

Mistral Large 3 675B Instruct 2512 NVFP4 is an Apache-2.0 licensed, instruct post-trained multimodal granular Mixture-of-Experts model. It has 675B total parameters with 41B active parameters, comprising a 673B-parameter MoE language model with 39B active and a 2.5B vision encoder. It is an NVFP4 post-training-activation quantized checkpoint derived from Mistral-Large-3-675B-Instruct-2512 and created with llm-compressor in collaboration with vLLM and Red Hat. The model card reports support for vision, multilingual use, system prompts, agentic capabilities with native function calling and JSON output, and a 256k context window. It is described as deployable on-premises in NVFP4 on a single node of H100s or A100s, while the FP8 version is described as deployable on a single node of B200s or H200s.