Model Anatomy See what's inside open-weight AI models

Can I run deepseek-v4-gguf on an NVIDIA H100?

No

Even its smallest file (DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2, 86.7 GB) is more than an NVIDIA H100 can hold (72.0 GB usable): it needs 2 of them.

Full analysis of deepseek-v4-gguf Everything that runs on NVIDIA H100

Files and context on NVIDIA H100

Each file of deepseek-v4-gguf (and of GGUF versions the site has analyzed), and how much context fits next to it, with the context cache in 16-bit or 8-bit.

FileSizeOn NVIDIA H100 Context (16-bit)Context (8-bit)
DeepSeek-V4-Pro-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-Instruct-imatrix-0813465 GBDoesn't fit——
DeepSeek-V4-Pro-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-Instruct465 GBDoesn't fit——
DeepSeek-V4-Pro-Q4K-Layers00-30458 GBDoesn't fit——
DeepSeek-V4-Pro-Q4K-Layers-31-output442 GBDoesn't fit——
DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-imatrix-0731165 GBDoesn't fit——
DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2165 GBDoesn't fit——
DeepSeek-V4-Flash-Vision-Exp-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out156 GBDoesn't fit——
MXFP4156 GBDoesn't fit——
DeepSeek-V4-Flash-Vision-Exp-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ897.6 GBDoesn't fit——
DeepSeek-V4-Flash-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-fixed-073197.6 GBDoesn't fit——
DeepSeek-V4-Flash-Layers37-42Q4KExperts-OtherExpertLayersIQ2XXSGateUp-Q2KDown-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-fixed97.6 GBDoesn't fit——
DeepSeek-V4-Flash-Vision-Exp-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ886.7 GBDoesn't fit——
DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-073186.7 GBDoesn't fit——
DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v286.7 GBDoesn't fit——

deepseek-v4-gguf on other hardware

Popular models that run on NVIDIA H100

Assumes 90% of the card's memory is usable and at least 4K tokens of context for one conversation, with the context cache in 16-bit unless noted. Fits leaves at least 20% of that memory for context; Tight leaves less.