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 64K context with q8_0 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
vision language 191text 1798image 2audio tts 21audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2093 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Step-3.7-FlashQ4_1201B116.67 GiB13.30 GiB131.00 GiB0.13 GiB21±22%
grok-2MoEQ3_K_L270B121.28 GiB8.50 GiB130.92 GiB0.21 GiB35±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB3.98 GiB130.86 GiB0.27 GiB98±37%
GLM-4.7-REAP-218B-A32BMoEQ4_0218B117.58 GiB12.22 GiB130.83 GiB0.30 GiB54±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.66 GiB130.80 GiB0.33 GiB120±37%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_NL236B123.50 GiB6.24 GiB130.77 GiB0.36 GiB72±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_NL236B123.50 GiB6.24 GiB130.77 GiB0.36 GiB72±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB3.98 GiB130.56 GiB0.57 GiB98±37%
MiniMax-M2.5MoEQ4_K_S229B121.10 GiB8.23 GiB130.32 GiB0.81 GiB79±37%
MiniMax-M2.1MoEI1-Q4_K_S229B121.10 GiB8.23 GiB130.32 GiB0.81 GiB79±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_S229B121.10 GiB8.23 GiB130.32 GiB0.81 GiB79±37%
MiniMax-M2.7MoEQ4_0229B120.93 GiB8.23 GiB130.15 GiB0.98 GiB79±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB3.98 GiB129.99 GiB1.14 GiB98±37%
MiniMax-M2MoEIQ4_NL229B120.37 GiB8.23 GiB129.59 GiB1.54 GiB79±37%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.03 GiB129.51 GiB1.62 GiB137±37%
step-3.5-flashQ4_1199B115.15 GiB13.30 GiB129.48 GiB1.65 GiB21±22%
Hermes-4-405BIQ2_XS406B111.15 GiB16.73 GiB129.17 GiB1.96 GiB21±22%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB16.73 GiB129.17 GiB1.96 GiB21±22%
Hy3MoEIQ3_XXS299B117.43 GiB10.63 GiB129.09 GiB2.04 GiB68±37%
dots.llm1.instMoEQ5_K_S143B95.08 GiB32.94 GiB129.05 GiB2.08 GiB36±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB7.17 GiB128.55 GiB2.58 GiB21±22%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.03 GiB128.36 GiB2.77 GiB138±37%
command-a-plus-05-2026-bf16MoEQ4_K_L219B126.06 GiB1.29 GiB128.35 GiB2.78 GiB92±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB2.24 GiB127.96 GiB3.17 GiB105±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB2.24 GiB127.96 GiB3.17 GiB105±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB2.24 GiB127.96 GiB3.17 GiB105±37%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB2.28 GiB127.74 GiB3.39 GiB110±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB2.28 GiB127.74 GiB3.39 GiB110±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB2.41 GiB127.31 GiB3.82 GiB125±37%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB12.22 GiB127.29 GiB3.84 GiB62±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB12.22 GiB127.29 GiB3.84 GiB62±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB12.22 GiB126.82 GiB4.31 GiB62±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
Trinity-Large-TrueBaseMoEI1-Q2_K_S399B122.88 GiB2.41 GiB126.31 GiB4.82 GiB126±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_M402B118.78 GiB6.38 GiB126.18 GiB4.95 GiB103±37%
Ornith-1.0-397BMoEIQ2_M397B123.99 GiB1.00 GiB126.04 GiB5.09 GiB135±37%
Llama-3_1-Nemotron-51B-InstructQ6_K_L51.5B39.83 GiB85.00 GiB125.97 GiB5.16 GiB22±22%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.80 GiB125.31 GiB5.82 GiB124±37%
Qwen3-235B-A22B-abliteratedMoEIQ4_XS235B117.97 GiB6.24 GiB125.24 GiB5.89 GiB75±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q6_K_L49.9B38.58 GiB85.00 GiB124.72 GiB6.41 GiB22±22%
Valkyrie-49B-v2.1Q6_K_L49.9B38.58 GiB85.00 GiB124.72 GiB6.41 GiB22±22%
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB6.24 GiB124.51 GiB6.62 GiB75±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB6.24 GiB124.51 GiB6.62 GiB75±37%
Llama-3_3-Nemotron-Super-49B-v1Q6_K49.9B38.11 GiB85.00 GiB124.25 GiB6.88 GiB22±22%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB6.24 GiB124.16 GiB6.97 GiB75±37%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB2.92 GiB123.57 GiB7.56 GiB99±37%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.80 GiB122.78 GiB8.35 GiB126±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB1.67 GiB122.60 GiB8.53 GiB111±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB12.22 GiB121.73 GiB9.40 GiB60±37%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB1.00 GiB121.60 GiB9.53 GiB119±37%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB1.00 GiB120.62 GiB10.51 GiB141±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB12.22 GiB120.40 GiB10.73 GiB64±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB2.41 GiB119.94 GiB11.19 GiB131±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.75 GiB119.57 GiB11.56 GiB129±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB6.38 GiB119.03 GiB12.10 GiB95±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB1.00 GiB118.46 GiB12.67 GiB129±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB1.21 GiB117.96 GiB13.17 GiB126±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB8.23 GiB117.12 GiB14.01 GiB70±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB6.11 GiB116.53 GiB14.60 GiB79±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 65,536 context with q8_0 KV cache, the largest being Step-3.7-Flash at Q4_1. 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.