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. 2094 of 2118 indexed models fit at 8K context with q4_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
text 1799vision language 191image 2audio tts 21audio asr 39video 16embedding 26

What fits at 8K context

largest quantization that fits, per model · 2094 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Trinity-Large-ThinkingMoEQ2_K399B129.66 GiB0.35 GiB131.04 GiB0.09 GiB144±37%
Ornith-1.0-397BMoEQ2_K397B129.92 GiB0.07 GiB131.04 GiB0.09 GiB141±37%
Nex-N2-ProMoEQ2_K397B129.92 GiB0.07 GiB131.04 GiB0.09 GiB141±37%
MiniMax-M2.1MoEQ4_K_L229B129.49 GiB0.54 GiB131.02 GiB0.11 GiB121±37%
Trinity-Large-TrueBaseMoEQ2_K399B129.44 GiB0.35 GiB130.82 GiB0.31 GiB144±37%
MiniMax-M2.7MoEQ4_K_M229B129.27 GiB0.54 GiB130.80 GiB0.33 GiB121±37%
Step-3.7-FlashUD-Q5_K_S201B128.59 GiB1.13 GiB130.75 GiB0.38 GiB21±22%
MiniMax-M2MoEQ4_K_M229B129.07 GiB0.54 GiB130.60 GiB0.53 GiB121±37%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB0.19 GiB130.49 GiB0.64 GiB96±37%
MiniMax-M2.5MoEQ4_K_M229B128.84 GiB0.54 GiB130.37 GiB0.76 GiB121±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_M229B128.84 GiB0.54 GiB130.37 GiB0.76 GiB121±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.04 GiB130.18 GiB0.95 GiB126±37%
GLM-4.7-REAP-218B-A32BMoEQ4_1218B128.16 GiB0.81 GiB130.01 GiB1.12 GiB81±37%
Hy3MoEIQ3_XS299B128.00 GiB0.70 GiB129.74 GiB1.39 GiB110±37%
Yi-1.5-34BF3234.4B128.11 GiB0.53 GiB129.72 GiB1.41 GiB21±22%
Qwen3-235B-A22B-Instruct-2507MoEQ4_K_S235B128.25 GiB0.41 GiB129.69 GiB1.44 GiB95±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_S235B128.25 GiB0.41 GiB129.69 GiB1.44 GiB95±37%
Hermes-3-Llama-3.1-405BIQ2_M406B127.28 GiB1.11 GiB129.67 GiB1.46 GiB21±22%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.02 GiB129.50 GiB1.63 GiB137±37%
step-3.5-flashQ5_K_S199B126.55 GiB1.13 GiB128.72 GiB2.41 GiB21±22%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.02 GiB128.34 GiB2.79 GiB139±37%
Solar-Open2-250BMoEIQ4_XS250B126.88 GiB0.42 GiB128.33 GiB2.80 GiB132±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB0.42 GiB127.36 GiB3.77 GiB160±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB0.26 GiB127.14 GiB3.99 GiB127±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB0.26 GiB126.83 GiB4.30 GiB127±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB0.26 GiB126.27 GiB4.86 GiB128±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB0.54 GiB125.99 GiB5.14 GiB107±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB0.41 GiB125.95 GiB5.18 GiB98±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB0.41 GiB125.95 GiB5.18 GiB98±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB0.41 GiB125.95 GiB5.18 GiB98±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB0.41 GiB125.95 GiB5.18 GiB98±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB0.15 GiB125.87 GiB5.26 GiB122±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB0.15 GiB125.87 GiB5.26 GiB122±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB0.15 GiB125.87 GiB5.26 GiB122±37%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB0.77 GiB125.66 GiB5.47 GiB22±22%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB0.77 GiB125.66 GiB5.47 GiB22±22%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB0.15 GiB125.61 GiB5.52 GiB129±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB0.15 GiB125.61 GiB5.52 GiB129±37%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB0.54 GiB124.94 GiB6.19 GiB95±37%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.05 GiB124.57 GiB6.56 GiB131±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB0.81 GiB124.17 GiB6.96 GiB106±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB0.81 GiB124.17 GiB6.96 GiB106±37%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
QwQ-32BBF1632.8B122.07 GiB0.56 GiB123.73 GiB7.40 GiB22±22%
GLM-4.6MoEQ2_K_L357B121.85 GiB0.81 GiB123.70 GiB7.43 GiB107±37%
dots.llm1.instMoEQ6_K143B120.06 GiB2.18 GiB123.26 GiB7.87 GiB101±37%
Behemoth-X-123B-v2Q8_0123B121.33 GiB0.77 GiB123.26 GiB7.87 GiB22±22%
Mistral-Large-Instruct-2411Q8_0123B121.33 GiB0.77 GiB123.26 GiB7.87 GiB22±22%
grok-2MoEQ3_K_L270B121.28 GiB0.56 GiB122.98 GiB8.15 GiB42±37%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.05 GiB122.03 GiB9.10 GiB134±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB0.47 GiB121.85 GiB9.28 GiB23±22%
GLM-4.6-REAP-268B-A32BMoEQ3_K_M269B119.90 GiB0.81 GiB121.75 GiB9.38 GiB96±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB0.15 GiB121.08 GiB10.05 GiB124±37%
GLM-4.6-Derestricted-v3MoEQ2_K_L357B119.21 GiB0.81 GiB121.06 GiB10.07 GiB109±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB0.19 GiB120.84 GiB10.29 GiB120±37%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.07 GiB120.67 GiB10.46 GiB127±37%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.07 GiB119.69 GiB11.44 GiB153±37%
Hermes-4-405BIQ2_S406B117.02 GiB1.11 GiB119.41 GiB11.72 GiB23±22%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.05 GiB118.87 GiB12.26 GiB137±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?
2094 of 2118 indexed open-weight models fit a H200 SXM at 8,192 context with q4_0 KV cache, the largest being Trinity-Large-Thinking at Q2_K. 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.