mistralai

Mistral Large 4

Mistral's largest natively multimodal MoE model, about 1T total and 52B active parameters, with 1M-token context support.Mistral Ecosystem Open-Weight License (modified MIT)Hugging Face

No server in stock right now has the 1273 GB this model needs for inference.

License restrictions

Operators of a Model-as-a-Service business need a separate commercial license from Mistral before any commercial use.

GPU memory it needs

From model cards and vendor documentation. Lower precision needs less memory, with slightly lower quality. Estimates are marked.

What you want to do	Precision	Minimum	Recommended	Setup	Basis
Run the model vLLM recipe minimum for this repo's FP8 block-scaled weights, about 128 GiB of weights per GPU on 8 GPUs; recipe base max-model-len 262,144.	FP8, 8-bit	1273 GB	2304 GB	8 GPUs: 2x GB300 NVL4 trays	Published
Run the model vLLM recipe minimum for separate repo mistralai/Mistral-Large-4-1T-A52B-NVFP4, validated on GB300, B300 and B200.	4-bit	733 GB	1152 GB	1x GB300 NVL4 tray (4 GPUs, vLLM recipe default)	Published