NVIDIA · datacenter

H100 PCIe 80GB

H100 PCIe 80GB has 80 GB of VRAM at 2000 GB/s — about 74.40 GiB usable after driver and compositor overhead. 2017 of 2118 indexed models fit at 128K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
80 GB
HBM2e
Bandwidth
2000 GB/s
5120-bit bus
Tensor FP16
756 TF
dense
TDP
350 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1728vision language 185image 2audio asr 39audio tts 21embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 2017 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Apertus-70B-Instruct-2509Q3_K_M70.6B33.10 GiB40.00 GiB74.28 GiB0.12 GiB16±22%
Qwen3-Coder-NextMoEQ6_K79.7B61.27 GiB12.00 GiB74.26 GiB0.14 GiB38±37%
Qwen3-Next-80B-A3B-ThinkingMoEQ6_K81.3B61.27 GiB12.00 GiB74.26 GiB0.14 GiB38±37%
Qwen3-Next-80B-A3B-InstructMoEQ6_K81.3B61.27 GiB12.00 GiB74.26 GiB0.14 GiB38±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ1_S310B58.20 GiB15.00 GiB74.25 GiB0.15 GiB32±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
HuatuoGPT-o1-72BIQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
MiroThinker-v1.0-72BI1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
EVA-Qwen2.5-72B-v0.2IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Qwen2.5-Math-72B-InstructIQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Qwen2.5-72B-InstructIQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Qwen2.5-72BI1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
magnum-v4-72bI1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
KAT-Dev-72B-ExpIQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Homer-v1.0-Qwen2.5-72BIQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Qwen2.5-VL-72B-InstructIQ3_M73.4B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
Chronos-Platinum-72BIQ3_M72.7B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
UI-TARS-72B-DPOIQ3_M73.4B33.07 GiB40.00 GiB74.20 GiB0.20 GiB16±22%
GLM-4.6VMoEIQ3_M108B50.11 GiB23.00 GiB74.14 GiB0.26 GiB22±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ1_M235B49.57 GiB23.50 GiB74.10 GiB0.30 GiB22±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ1_M235B49.57 GiB23.50 GiB74.10 GiB0.30 GiB22±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ1_M235B49.49 GiB23.50 GiB74.02 GiB0.38 GiB22±37%
Assistant_Pepe_70BQ3_K_M70.6B32.89 GiB40.00 GiB74.02 GiB0.38 GiB16±22%
GLM-4.5-Air-REAP-82B-A12BMoEQ4_181.9B49.92 GiB23.00 GiB73.95 GiB0.45 GiB22±37%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q4_1117B68.42 GiB4.53 GiB73.94 GiB0.46 GiB60±37%
gpt-oss-120b-abliteratedMoEI1-Q4_1117B68.42 GiB4.53 GiB73.94 GiB0.46 GiB60±37%
Llama-2-7b-chat-hfQ5_K_M6.7B8.91 GiB64.00 GiB73.93 GiB0.47 GiB16±22%
Janus-Pro-7BF167.4B12.88 GiB60.00 GiB73.90 GiB0.50 GiB16±22%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ3_K_L124B61.83 GiB11.00 GiB73.83 GiB0.57 GiB37±37%
CalmeRys-78B-Orpo-v0.1I1-Q2_K78.0B29.66 GiB43.00 GiB73.79 GiB0.61 GiB16±22%
calme-2.3-rys-78bQ2_K78.0B29.66 GiB43.00 GiB73.79 GiB0.61 GiB16±22%
c4ai-command-r-plus-08-2024IQ3_XS104B40.61 GiB32.00 GiB73.79 GiB0.61 GiB16±22%
GLM-4.5-Air-DerestrictedMoEQ3_K_S110B49.75 GiB23.00 GiB73.78 GiB0.62 GiB22±37%
GLM-4.5-AirMoEQ3_K_S110B49.75 GiB23.00 GiB73.78 GiB0.62 GiB22±37%
Qwen3.5-122B-A10BMoEQ4_K_S125B69.66 GiB3.00 GiB73.69 GiB0.71 GiB68±37%
Devstral-2-123B-Instruct-2512UD-IQ1_M125B28.47 GiB44.00 GiB73.62 GiB0.78 GiB16±22%
Mistral-MOE-4X7B-Dark-MultiVerse-Uncensored-Enhanced32-24BMoEQ6_K24.2B56.49 GiB16.00 GiB73.53 GiB0.87 GiB9±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_XS236B63.99 GiB8.44 GiB73.47 GiB0.93 GiB44±37%
DeepSeek-V2.5MoEIQ2_XS236B63.99 GiB8.44 GiB73.47 GiB0.93 GiB44±37%
DeepSeek-Coder-V2-InstructMoEIQ2_XS236B63.99 GiB8.44 GiB73.47 GiB0.93 GiB44±37%
Kimi-Dev-72BQ3_K_S72.7B32.12 GiB40.00 GiB73.25 GiB1.15 GiB16±22%
Chuluun-Qwen2.5-72B-v0.01Q3_K_S72.7B32.12 GiB40.00 GiB73.25 GiB1.15 GiB16±22%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_M123B69.11 GiB3.00 GiB73.14 GiB1.26 GiB69±37%
Qwen3-72B-SynthesisQ3_K_S72.7B31.95 GiB40.00 GiB73.08 GiB1.32 GiB16±22%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ8_035.1B69.57 GiB2.50 GiB73.08 GiB1.32 GiB72±37%
Meta-Llama-3-70B-InstructQ3_K_M70.6B31.92 GiB40.00 GiB73.05 GiB1.35 GiB16±22%
Maenad-70BI1-Q3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
calme-2.4-llama3-70bQ3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
calme-2.2-llama3-70bQ3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
Rombos-LLM-70b-Llama-3.3I1-Q3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
L3.3-Electra-R1-70bI1-Q3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
L3.3-70B-Magnum-v4-SEQ3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
Llama-3.3_70_b_uncensored_continuedI1-Q3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
Llama-3.3-70B-Instruct-abliteratedI1-Q3_K_M70.6B31.91 GiB40.00 GiB73.04 GiB1.36 GiB16±22%
●From the file◐Predictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Questions people ask

What AI models can a H100 PCIe 80GB run?
2017 of 2118 indexed open-weight models fit a H100 PCIe 80GB at 131,072 context with f16 KV cache, the largest being Apertus-70B-Instruct-2509 at Q3_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a H100 PCIe 80GB actually have?
Its nameplate is 80 GB, but about 74.40 GiB is available to a model once driver and compositor overhead is accounted for.
Is a H100 PCIe 80GB fast for local AI?
Its memory bandwidth is 2000 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.