NVIDIA · datacenter

H200 SXM

H200 SXM has 141 GB of VRAM at 4800 GB/s — about 131.13 GiB usable after driver and compositor overhead. 2093 of 2118 indexed models fit at 32K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
141 GB
HBM3e
Bandwidth
4800 GB/s
6144-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 1798vision language 191image 2audio tts 21audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2093 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiniMax-M2.1MoEQ4_0229B122.32 GiB7.75 GiB131.05 GiB0.08 GiB80±37%
MiniMax-M2MoEQ4_0229B122.32 GiB7.75 GiB131.05 GiB0.08 GiB80±37%
GLM-4.6-Derestricted-v3MoEQ2_K357B118.51 GiB11.50 GiB131.05 GiB0.08 GiB63±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_0236B123.99 GiB5.88 GiB130.90 GiB0.23 GiB73±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_0236B123.99 GiB5.88 GiB130.90 GiB0.23 GiB73±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_0235B123.99 GiB5.88 GiB130.90 GiB0.23 GiB73±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_NL235B123.96 GiB5.88 GiB130.86 GiB0.27 GiB73±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ4_NL235B123.96 GiB5.88 GiB130.86 GiB0.27 GiB73±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.63 GiB130.76 GiB0.37 GiB120±37%
Step-3.7-FlashQ4_1201B116.67 GiB13.03 GiB130.73 GiB0.40 GiB21±22%
MiniMax-M2.7MoEUD-Q4_K_S229B121.97 GiB7.75 GiB130.70 GiB0.43 GiB81±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB3.75 GiB130.63 GiB0.50 GiB99±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q8_049.9B49.36 GiB80.00 GiB130.50 GiB0.63 GiB21±22%
Valkyrie-49B-v2.1Q8_049.9B49.36 GiB80.00 GiB130.50 GiB0.63 GiB21±22%
Llama-3_3-Nemotron-Super-49B-v1Q8_049.9B49.36 GiB80.00 GiB130.50 GiB0.63 GiB21±22%
grok-2MoEQ3_K_L270B121.28 GiB8.00 GiB130.42 GiB0.71 GiB35±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB3.75 GiB130.32 GiB0.81 GiB99±37%
GLM-4.7-REAP-218B-A32BMoEQ4_0218B117.58 GiB11.50 GiB130.12 GiB1.01 GiB55±37%
MiniMax-M2.5MoEQ4_K_S229B121.10 GiB7.75 GiB129.84 GiB1.29 GiB81±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_S229B121.10 GiB7.75 GiB129.84 GiB1.29 GiB81±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB3.75 GiB129.76 GiB1.37 GiB100±37%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.06 GiB129.54 GiB1.59 GiB137±37%
step-3.5-flashQ4_1199B115.15 GiB13.03 GiB129.21 GiB1.92 GiB21±22%
command-a-plus-05-2026-bf16MoEQ4_K_L219B126.06 GiB1.42 GiB128.48 GiB2.65 GiB91±37%
Hy3MoEIQ3_XXS299B117.43 GiB10.00 GiB128.47 GiB2.66 GiB70±37%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.06 GiB128.39 GiB2.74 GiB138±37%
Hermes-4-405BIQ2_XS406B111.15 GiB15.75 GiB128.18 GiB2.95 GiB22±22%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB15.75 GiB128.18 GiB2.95 GiB22±22%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB6.75 GiB128.12 GiB3.01 GiB22±22%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB2.11 GiB127.83 GiB3.30 GiB106±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB2.11 GiB127.83 GiB3.30 GiB106±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB2.11 GiB127.83 GiB3.30 GiB106±37%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB2.14 GiB127.61 GiB3.52 GiB111±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB2.14 GiB127.61 GiB3.52 GiB111±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB2.67 GiB127.57 GiB3.56 GiB123±37%
dots.llm1.instMoEQ5_K_S143B95.08 GiB31.00 GiB127.11 GiB4.02 GiB38±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB11.50 GiB126.57 GiB4.56 GiB64±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB11.50 GiB126.57 GiB4.56 GiB64±37%
Trinity-Large-TrueBaseMoEI1-Q2_K_S399B122.88 GiB2.67 GiB126.57 GiB4.56 GiB124±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB11.50 GiB126.10 GiB5.03 GiB64±37%
Ornith-1.0-397BMoEIQ2_M397B123.99 GiB0.94 GiB125.98 GiB5.15 GiB136±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_M402B118.78 GiB6.00 GiB125.81 GiB5.32 GiB106±37%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.75 GiB125.27 GiB5.86 GiB124±37%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB5.88 GiB123.79 GiB7.34 GiB76±37%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB2.75 GiB123.40 GiB7.73 GiB101±37%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.75 GiB122.73 GiB8.40 GiB126±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB1.64 GiB122.57 GiB8.56 GiB111±37%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.94 GiB121.54 GiB9.59 GiB119±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB11.50 GiB121.01 GiB10.12 GiB61±37%
Llama-3_1-Nemotron-51B-InstructQ6_K_L51.5B39.83 GiB80.00 GiB120.97 GiB10.16 GiB23±22%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.94 GiB120.56 GiB10.57 GiB141±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB2.67 GiB120.20 GiB10.93 GiB128±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.70 GiB119.52 GiB11.61 GiB130±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB6.00 GiB118.66 GiB12.47 GiB97±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.94 GiB118.41 GiB12.72 GiB130±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB1.15 GiB117.91 GiB13.22 GiB127±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB7.75 GiB116.63 GiB14.50 GiB72±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB5.75 GiB116.17 GiB14.96 GiB80±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 H200 SXM run?
2093 of 2118 indexed open-weight models fit a H200 SXM at 32,768 context with f16 KV cache, the largest being MiniMax-M2.1 at Q4_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a H200 SXM actually have?
Its nameplate is 141 GB, but about 131.13 GiB is available to a model once driver and compositor overhead is accounted for.
Is a H200 SXM fast for local AI?
Its memory bandwidth is 4800 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.