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 q8_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◐
GLM-4.5MoEIQ1_S358B71.81 GiB1.53 GiB74.38 GiB0.02 GiB112±37%
Qwen3.5-40B-Claude-4.5-Opus-High-Reasoning-Thinking-uncensored-hereticBF1639.5B72.80 GiB0.40 GiB74.26 GiB0.14 GiB26±22%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-Q2_K_S229B72.23 GiB1.03 GiB74.25 GiB0.15 GiB135±37%
MiniMax-M2.1MoEI1-Q2_K_S229B72.23 GiB1.03 GiB74.25 GiB0.15 GiB135±37%
MiniMax-M2.5MoEI1-Q2_K_S229B72.23 GiB1.03 GiB74.25 GiB0.15 GiB135±37%
Mistral-MOE-4X7B-Dark-MultiVerse-Uncensored-Enhanced32-24BMoEQ8_024.2B72.66 GiB0.53 GiB74.23 GiB0.17 GiB15±37%
Behemoth-X-123B-v2Q4_1123B71.45 GiB1.46 GiB74.06 GiB0.34 GiB26±22%
GLM-4.7MoEIQ1_S358B71.50 GiB1.53 GiB74.06 GiB0.34 GiB112±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ4_K_S124B72.68 GiB0.37 GiB74.04 GiB0.36 GiB132±37%
Qwen3.5-122B-A10BMoEUD-Q4_K_M125B72.89 GiB0.10 GiB74.01 GiB0.39 GiB152±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q4_1125B72.82 GiB0.10 GiB73.95 GiB0.45 GiB152±37%
MiMo-V2-FlashMoEKV unresolvedUD-TQ1_0310B72.36 GiB0.50 GiB73.90 GiB0.50 GiB145±37%
dots.llm1.instMoEQ3_K_M143B68.74 GiB4.12 GiB73.88 GiB0.52 GiB92±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_1121B70.61 GiB2.14 GiB73.78 GiB0.62 GiB26±22%
DeepSeek-V4-Flash-162BMoEKV unresolvedQ3_K_M92.2B72.27 GiB0.36 GiB73.68 GiB0.72 GiB135±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_M235B71.85 GiB0.78 GiB73.67 GiB0.73 GiB111±37%
GLM-4.6-Derestricted-v3MoEIQ1_S357B71.07 GiB1.53 GiB73.63 GiB0.77 GiB113±37%
GLM-4.6MoEIQ1_S357B71.07 GiB1.53 GiB73.63 GiB0.77 GiB113±37%
ERNIE-4.5-300B-A47B-PTUD-TQ1_0300B71.48 GiB0.90 GiB73.50 GiB0.90 GiB26±22%
Devstral-2-123B-Instruct-2512Q4_K_L125B70.87 GiB1.46 GiB73.49 GiB0.91 GiB26±22%
GLM-4.7-REAP-218B-A32BMoEUD-IQ2_M218B70.78 GiB1.53 GiB73.35 GiB1.05 GiB92±37%
MiniMax-M2.7MoEIQ2_M229B71.17 GiB1.03 GiB73.18 GiB1.22 GiB137±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ5_K_M109B71.29 GiB0.80 GiB73.11 GiB1.29 GiB111±37%
Mixtral-8x22B-Instruct-v0.1MoEIQ4_XS141B71.12 GiB0.93 GiB73.11 GiB1.29 GiB49±37%
Mixtral-8x22B-v0.1MoEIQ4_XS141B71.11 GiB0.93 GiB73.10 GiB1.30 GiB49±37%
Mixtral-8x22B-v0.1MoEIQ4_XS141B71.11 GiB0.93 GiB73.10 GiB1.30 GiB49±37%
Qwen3.5-REAP-262B-A17BMoEIQ2_XS262B71.81 GiB0.12 GiB72.99 GiB1.41 GiB155±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_M236B71.64 GiB0.28 GiB72.96 GiB1.44 GiB142±37%
DeepSeek-V2.5MoEIQ2_M236B71.64 GiB0.28 GiB72.96 GiB1.44 GiB142±37%
DeepSeek-Coder-V2-InstructMoEIQ2_M236B71.64 GiB0.28 GiB72.96 GiB1.44 GiB142±37%
GLM-4.5-Air-REAP-82B-A12BMoEQ6_K81.9B70.98 GiB0.76 GiB72.77 GiB1.63 GiB98±37%
command-a-plus-05-2026-bf16MoEIQ2_M219B71.32 GiB0.36 GiB72.68 GiB1.72 GiB116±37%
GLM-4.6VMoEQ5_K_S108B70.79 GiB0.76 GiB72.58 GiB1.82 GiB112±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_1123B71.35 GiB0.10 GiB72.48 GiB1.92 GiB154±37%
Apertus-70B-Instruct-2509Q8_070.6B69.87 GiB1.33 GiB72.38 GiB2.02 GiB27±22%
Mistral-Medium-3.5-128BI1-Q4_K_M128B69.75 GiB1.46 GiB72.37 GiB2.03 GiB27±22%
XORTRON-NXTXPRTXXLI1-Q4_K_M128B69.75 GiB1.46 GiB72.37 GiB2.03 GiB27±22%
GLM-4.5VMoEI1-Q5_K_S108B70.53 GiB0.76 GiB72.32 GiB2.08 GiB113±37%
Meta-Llama-3-70B-InstructQ8_070.6B69.83 GiB1.33 GiB72.29 GiB2.11 GiB27±22%
calme-2.4-llama3-70bQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
calme-2.2-llama3-70bQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
L3.3-Electra-R1-70bQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
L3.3-70B-Magnum-v4-SEQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Hermes-4-70BQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Llama-3.3_70_b_uncensored_continuedQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
grok-oss-Revenant-70BQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Llama-3.1-70BQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Llama-3.3-70B-InstructQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Llama-3.1-Nemotron-70B-Instruct-HFQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Llama-3.3-70B-Instruct-abliteratedQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Hermes-4-70B-hereticQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Anubis-70B-v1.2Q8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
L3.3-70B-Euryale-v2.3Q8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Rombos-LLM-70b-Llama-3.3Q8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
llama-3-firefunction-v2Q8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
DeepSeek-R1-Distill-Llama-70B-hereticQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
DeepSeek-R1-Distill-Llama-70BQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
DeepSeek-R1-Distill-Llama-70B-abliteratedQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Legion-V2.1-LLaMa-70BQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 GiB27±22%
Tess-R1-Limerick-Llama-3.1-70BQ8_070.6B69.83 GiB1.33 GiB72.28 GiB2.12 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 q8_0 KV cache, the largest being GLM-4.5 at IQ1_S. 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.