Mistral Large 3 675B Instruct 2512 NVFP4
Mistral Large 3 675B Instruct 2512 NVFP4 is an Apache-2.0 licensed, multimodal granular Mixture-of-Experts instruct model from Mistral AI. It has 675B total parameters with 41B active parameters, including a 2.5B vision encoder, and a 256k context window. It supports multilingual text, vision, system prompts, agentic capabilities, native function calling, and JSON output. It is a post-training-activation quantized NVFP4 version of Mistral-Large-3-675B-Instruct-2512, created with llm-compressor. The repository lists vLLM as its library and includes tags for mistral-common and compressed-tensors. The model card states deployment on-premises in NVFP4 on a single node of H100s or A100s and lists use cases such as long document understanding, assistants, agentic/tool use, enterprise knowledge work, and coding assistance.
At a glance
- Architecture
- unknown
- License
- apache-2.0
- Software
- vLLM
Will it run on your machine?
Save your machine to see a personalized rating and its reasoning.
Add my hardwareBest for
- Long document understanding
- Daily-driver AI assistants
- Agentic and tool-use applications
- Enterprise knowledge work
- General coding assistance
Tradeoffs
- The source lists deployment only in NVFP4 on a single node of H100s or A100s; other hardware configurations are not stated.
- The source does not specify exact numerical hardware requirements, performance benchmarks, or whether the model was tested.
- The source does not state additional licensing terms beyond apache-2.0 or compatibility with specific software versions.
- The source indicates the model card uses vllm and inference: false but does not provide detailed serving instructions.
- The source does not provide limitations, safety evaluations, or out-of-scope uses beyond the listed capabilities.
Ways to run it
vLLM · Linux
Repository-specific example documented in the publisher model card; confirm dependencies and platform in the source.
vllm serve mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4 --max-model-len 262144 --tensor-parallel-size 8 --tokenizer_mode mistral --config_format mistral --load_format mistral --enable-auto-tool-choice --tool-call-parser mistralKeep exploring
Bonsai 2 · 27B
A compressed 27B-class reasoning model with publisher-provided GGUF packs and a dedicated llama.cpp fork for CUDA, Metal, and CPU.
DeepSeek V4.1 Flash
A multimodal reasoning model with a compressed key-value cache, published as open weights.
Qwen3.8-Flash-Next
An experimental open-weight multimodal model with sparse attention and 262K native context.