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. 2071 of 2118 indexed models fit at 16K context with f16 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
vision language 186text 1781image 2audio asr 39audio tts 21video 16embedding 26

What fits at 16K context

largest quantization that fits, per model · 2071 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Step-3.7-FlashQ2_K201B66.34 GiB7.03 GiB74.39 GiB0.01 GiB26±22%
Mistral-Medium-3.5-128BQ4_0128B67.72 GiB5.50 GiB74.38 GiB0.02 GiB26±22%
Qwen3.5-122B-A10BMoEUD-Q4_K_M125B72.89 GiB0.38 GiB74.29 GiB0.11 GiB146±37%
Llama-3_1-Nemotron-51B-InstructQ5_K_S51.5B33.12 GiB40.00 GiB74.26 GiB0.14 GiB26±22%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q4_1125B72.82 GiB0.38 GiB74.23 GiB0.17 GiB146±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q5_K_M49.9B32.96 GiB40.00 GiB74.10 GiB0.30 GiB26±22%
Valkyrie-49B-v2.1I1-Q5_K_M49.9B32.96 GiB40.00 GiB74.10 GiB0.30 GiB26±22%
Llama-3_3-Nemotron-Super-49B-v1Q5_K_M49.9B32.96 GiB40.00 GiB74.10 GiB0.30 GiB26±22%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ4_XS139B69.15 GiB3.88 GiB74.01 GiB0.39 GiB91±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ4_XS139B69.15 GiB3.88 GiB74.01 GiB0.39 GiB91±37%
GLM-4.7-REAP-218B-A32BMoEUD-IQ2_XXS218B67.13 GiB5.75 GiB73.92 GiB0.48 GiB70±37%
MiniMax-M2.5MoEUD-IQ2_XXS229B69.03 GiB3.88 GiB73.89 GiB0.51 GiB103±37%
MiniMax-M2.1MoEUD-IQ2_XXS229B68.98 GiB3.88 GiB73.84 GiB0.56 GiB103±37%
MiniMax-M2MoEUD-IQ2_XXS229B68.92 GiB3.88 GiB73.78 GiB0.62 GiB103±37%
c4ai-command-r-plus-08-2024Q5_K_M104B68.57 GiB4.00 GiB73.75 GiB0.65 GiB26±22%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ4_1124B71.37 GiB1.38 GiB73.74 GiB0.66 GiB120±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_M236B71.64 GiB1.05 GiB73.73 GiB0.67 GiB129±37%
DeepSeek-V2.5MoEIQ2_M236B71.64 GiB1.05 GiB73.73 GiB0.67 GiB129±37%
DeepSeek-Coder-V2-InstructMoEIQ2_M236B71.64 GiB1.05 GiB73.73 GiB0.67 GiB129±37%
Qwen3.5-REAP-262B-A17BMoEIQ2_XS262B71.81 GiB0.47 GiB73.33 GiB1.07 GiB148±37%
step-3.5-flashQ2_K_L199B65.26 GiB7.03 GiB73.32 GiB1.08 GiB26±22%
command-a-plus-05-2026-bf16MoEIQ2_M219B71.32 GiB0.92 GiB73.25 GiB1.15 GiB110±37%
dots.llm1.instMoEQ2_K_L143B56.67 GiB15.50 GiB73.19 GiB1.21 GiB50±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ5_K_S109B69.16 GiB3.00 GiB73.18 GiB1.22 GiB93±37%
Devstral-2-123B-Instruct-2512Q4_K_S125B66.36 GiB5.50 GiB73.01 GiB1.39 GiB26±22%
XORTRON-NXTXPRTXXLI1-Q4_K_S128B66.36 GiB5.50 GiB73.01 GiB1.39 GiB26±22%
GLM-4.5-Air-DerestrictedMoEQ4_K_L110B68.88 GiB2.88 GiB72.79 GiB1.61 GiB94±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_1123B71.35 GiB0.38 GiB72.76 GiB1.64 GiB149±37%
Seed-OSS-36B-InstructBF1636.2B67.35 GiB4.00 GiB72.44 GiB1.96 GiB27±22%
Hermes-4.3-36BBF1636.2B67.35 GiB4.00 GiB72.44 GiB1.96 GiB27±22%
GLM-4.5-AirMoEQ4_K_M110B68.45 GiB2.88 GiB72.36 GiB2.04 GiB95±37%
Mixtral-8x22B-Instruct-v0.1MoEQ3_K_L141B67.60 GiB3.50 GiB72.16 GiB2.24 GiB46±37%
Mixtral-8x22B-v0.1MoEQ3_K_L141B67.60 GiB3.50 GiB72.16 GiB2.24 GiB46±37%
Mixtral-8x22B-v0.1MoEQ3_K_L141B67.60 GiB3.50 GiB72.16 GiB2.24 GiB46±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ4_K_M123B70.63 GiB0.38 GiB72.04 GiB2.36 GiB150±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_K_S121B63.83 GiB7.03 GiB71.89 GiB2.51 GiB27±22%
Behemoth-X-123B-v2Q4_K_S123B64.79 GiB5.50 GiB71.45 GiB2.95 GiB27±22%
Mistral-Large-Instruct-2411Q4_K_S123B64.79 GiB5.50 GiB71.45 GiB2.95 GiB27±22%
MiMo-V2-FlashMoEKV unresolvedIQ2_XXS310B68.47 GiB1.88 GiB71.40 GiB3.00 GiB127±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEQ8_035.1B69.57 GiB0.31 GiB70.89 GiB3.51 GiB153±37%
Laguna-S-2.1MoEQ4_1118B68.96 GiB0.89 GiB70.87 GiB3.53 GiB134±37%
GLM-4.6VMoEQ4_K_L108B66.89 GiB2.88 GiB70.79 GiB3.61 GiB96±37%
CalmeRys-78B-Orpo-v0.1Q6_K78.0B64.27 GiB5.38 GiB70.78 GiB3.62 GiB27±22%
calme-2.3-rys-78bQ6_K78.0B64.27 GiB5.38 GiB70.78 GiB3.62 GiB27±22%
MiniMax-M2.7MoEUD-IQ2_M229B65.32 GiB3.88 GiB70.18 GiB4.22 GiB106±37%
Mistral-Small-4-119B-2603MoEUD-Q4_K_M119B68.70 GiB0.35 GiB70.08 GiB4.32 GiB154±37%
HunyuanImage-2.1Q6_K17.5B68.97 GiB0.00 GiB70.02 GiB4.38 GiB28±22%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q4_1117B68.42 GiB0.59 GiB70.00 GiB4.40 GiB150±37%
gpt-oss-120b-abliteratedMoEI1-Q4_1117B68.42 GiB0.59 GiB70.00 GiB4.40 GiB150±37%
Qwen2.5-Coder-32B-InstructQ8_032.8B64.86 GiB4.00 GiB69.96 GiB4.44 GiB28±22%
MiMo-V2.5MoEKV unresolvedIQ1_M311B67.01 GiB1.88 GiB69.93 GiB4.47 GiB129±37%
GLM-4.5VMoEI1-Q4_K_M108B65.61 GiB2.88 GiB69.51 GiB4.89 GiB97±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_S235B65.40 GiB2.94 GiB69.37 GiB5.03 GiB97±37%
gpt-oss-20b-hereticMoEIQ4_NL20.9B67.58 GiB0.39 GiB68.96 GiB5.44 GiB84±37%
grok-2MoEIQ2_XXS270B63.81 GiB4.00 GiB68.95 GiB5.45 GiB47±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ1_M236B64.90 GiB2.94 GiB68.87 GiB5.53 GiB98±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ1_M236B64.83 GiB2.94 GiB68.80 GiB5.60 GiB98±37%
Qwen3.5-88BMoEI1-Q6_K87.7B67.08 GiB0.38 GiB68.48 GiB5.92 GiB140±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ2_S229B63.36 GiB3.88 GiB68.22 GiB6.18 GiB108±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEBF1635.1B66.19 GiB0.31 GiB67.50 GiB6.90 GiB160±37%
●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?
2071 of 2118 indexed open-weight models fit a H100 SXM 80GB at 16,384 context with f16 KV cache, the largest being Step-3.7-Flash at Q2_K. 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.