NVIDIA · datacenter

H100 SXM 80GB

H100 SXM 80GB has 80 GB of VRAM at 3350 GB/s — about 74.40 GiB usable after driver and compositor overhead. 2076 of 2118 indexed models fit at 8K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
80 GB
HBM3
Bandwidth
3350 GB/s
5120-bit bus
Tensor FP16
989 TF
dense
TDP
700 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1786vision language 186image 2audio asr 39audio tts 21video 16embedding 26

What fits at 8K context

largest quantization that fits, per model · 2076 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiniMax-M2.5MoEUD-IQ2_M229B72.83 GiB0.54 GiB74.36 GiB0.04 GiB143±37%
MiniMax-M2.1MoEUD-IQ2_M229B72.78 GiB0.54 GiB74.31 GiB0.09 GiB143±37%
MiniMax-M2MoEUD-IQ2_M229B72.72 GiB0.54 GiB74.25 GiB0.15 GiB143±37%
grok-2MoEIQ2_XS270B72.40 GiB0.56 GiB74.10 GiB0.30 GiB49±37%
Qwen3.5-40B-Claude-4.5-Opus-High-Reasoning-Thinking-uncensored-hereticBF1639.5B72.80 GiB0.21 GiB74.07 GiB0.33 GiB26±22%
Qwen2.5-72B-InstructQ8_072.7B72.21 GiB0.70 GiB74.04 GiB0.36 GiB26±22%
Mistral-MOE-4X7B-Dark-MultiVerse-Uncensored-Enhanced32-24BMoEQ8_024.2B72.66 GiB0.28 GiB73.98 GiB0.42 GiB15±37%
Qwen3.5-122B-A10BMoEUD-Q4_K_M125B72.89 GiB0.05 GiB73.97 GiB0.43 GiB153±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q4_1125B72.82 GiB0.05 GiB73.90 GiB0.50 GiB153±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ4_K_S124B72.68 GiB0.19 GiB73.87 GiB0.53 GiB135±37%
HuatuoGPT-o1-72BQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Qwen2.5-72B-Instruct-abliteratedQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
MiroThinker-v1.0-72BQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Qwen2.5-Math-72B-InstructQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Qwen2.5-72BQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Rombo-LLM-V3.0-Qwen-72bQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
EVA-Qwen2.5-72B-v0.2Q8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Kimi-Dev-72BQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Chuluun-Qwen2.5-72B-v0.01Q8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
magnum-v4-72bQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
KAT-Dev-72B-ExpQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Homer-v1.0-Qwen2.5-72BQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Qwen2.5-VL-72B-InstructQ8_073.4B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Tower-Plus-72B-ultra-uncensored-hereticQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
Chronos-Platinum-72BQ8_072.7B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
UI-TARS-72B-DPOQ8_073.4B71.96 GiB0.70 GiB73.79 GiB0.61 GiB26±22%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-Q2_K_S229B72.23 GiB0.54 GiB73.76 GiB0.64 GiB144±37%
MiMo-V2-FlashMoEKV unresolvedUD-TQ1_0310B72.36 GiB0.26 GiB73.67 GiB0.73 GiB149±37%
GLM-4.5MoEIQ1_S358B71.81 GiB0.81 GiB73.66 GiB0.74 GiB120±37%
DeepSeek-V4-Flash-162BMoEKV unresolvedQ3_K_M92.2B72.27 GiB0.19 GiB73.51 GiB0.89 GiB138±37%
Behemoth-X-123B-v2Q4_1123B71.45 GiB0.77 GiB73.38 GiB1.02 GiB26±22%
GLM-4.7MoEIQ1_S358B71.50 GiB0.81 GiB73.34 GiB1.06 GiB121±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_M235B71.85 GiB0.41 GiB73.30 GiB1.10 GiB115±37%
ERNIE-4.5-300B-A47B-PTUD-TQ1_0300B71.48 GiB0.47 GiB73.08 GiB1.32 GiB26±22%
Qwen3.5-REAP-262B-A17BMoEIQ2_XS262B71.81 GiB0.07 GiB72.93 GiB1.47 GiB156±37%
GLM-4.6-Derestricted-v3MoEIQ1_S357B71.07 GiB0.81 GiB72.92 GiB1.48 GiB122±37%
GLM-4.6MoEIQ1_S357B71.07 GiB0.81 GiB72.92 GiB1.48 GiB122±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_M236B71.64 GiB0.15 GiB72.83 GiB1.57 GiB144±37%
DeepSeek-V2.5MoEIQ2_M236B71.64 GiB0.15 GiB72.83 GiB1.57 GiB144±37%
DeepSeek-Coder-V2-InstructMoEIQ2_M236B71.64 GiB0.15 GiB72.83 GiB1.57 GiB144±37%
Devstral-2-123B-Instruct-2512Q4_K_L125B70.87 GiB0.77 GiB72.80 GiB1.60 GiB27±22%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_1121B70.61 GiB1.13 GiB72.77 GiB1.63 GiB27±22%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ5_K_M109B71.29 GiB0.42 GiB72.74 GiB1.66 GiB115±37%
MiniMax-M2.7MoEIQ2_M229B71.17 GiB0.54 GiB72.70 GiB1.70 GiB146±37%
Mixtral-8x22B-Instruct-v0.1MoEIQ4_XS141B71.12 GiB0.49 GiB72.67 GiB1.73 GiB50±37%
Mixtral-8x22B-v0.1MoEIQ4_XS141B71.11 GiB0.49 GiB72.66 GiB1.74 GiB50±37%
Mixtral-8x22B-v0.1MoEIQ4_XS141B71.11 GiB0.49 GiB72.66 GiB1.74 GiB50±37%
GLM-4.7-REAP-218B-A32BMoEUD-IQ2_M218B70.78 GiB0.81 GiB72.63 GiB1.77 GiB98±37%
command-a-plus-05-2026-bf16MoEIQ2_M219B71.32 GiB0.19 GiB72.52 GiB1.88 GiB118±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_1123B71.35 GiB0.05 GiB72.43 GiB1.97 GiB155±37%
GLM-4.5-Air-REAP-82B-A12BMoEQ6_K81.9B70.98 GiB0.40 GiB72.41 GiB1.99 GiB101±37%
GLM-4.6VMoEQ5_K_S108B70.79 GiB0.40 GiB72.22 GiB2.18 GiB116±37%
GLM-4.5VMoEI1-Q5_K_S108B70.53 GiB0.40 GiB71.96 GiB2.44 GiB117±37%
dots.llm1.instMoEQ3_K_M143B68.74 GiB2.18 GiB71.94 GiB2.46 GiB110±37%
Apertus-70B-Instruct-2509Q8_070.6B69.87 GiB0.70 GiB71.76 GiB2.64 GiB27±22%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ4_K_M123B70.63 GiB0.05 GiB71.71 GiB2.69 GiB157±37%
Mistral-Medium-3.5-128BI1-Q4_K_M128B69.75 GiB0.77 GiB71.68 GiB2.72 GiB27±22%
XORTRON-NXTXPRTXXLI1-Q4_K_M128B69.75 GiB0.77 GiB71.68 GiB2.72 GiB27±22%
Meta-Llama-3-70B-InstructQ8_070.6B69.83 GiB0.70 GiB71.66 GiB2.74 GiB27±22%
calme-2.4-llama3-70bQ8_070.6B69.83 GiB0.70 GiB71.65 GiB2.75 GiB27±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 SXM 80GB run?
2076 of 2118 indexed open-weight models fit a H100 SXM 80GB at 8,192 context with q4_0 KV cache, the largest being MiniMax-M2.5 at UD-IQ2_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a H100 SXM 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 SXM 80GB fast for local AI?
Its memory bandwidth is 3350 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.