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 16K 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
text 1799vision language 191image 2audio tts 21audio asr 39video 16embedding 26

What fits at 16K context

largest quantization that fits, per model · 2094 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-235B-A22B-Instruct-2507MoEQ4_K_S235B128.25 GiB1.56 GiB130.84 GiB0.29 GiB89±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_S235B128.25 GiB1.56 GiB130.84 GiB0.29 GiB89±37%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB0.49 GiB130.79 GiB0.34 GiB94±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.17 GiB130.30 GiB0.83 GiB125±37%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.03 GiB129.51 GiB1.62 GiB137±37%
Solar-Open2-250BMoEIQ4_XS250B126.88 GiB1.59 GiB129.50 GiB1.63 GiB120±37%
dots.llm1.instMoEQ6_K143B120.06 GiB8.23 GiB129.32 GiB1.81 GiB73±37%
GLM-4.7-REAP-218B-A32BMoEQ4_K_M218B124.65 GiB3.05 GiB128.74 GiB2.39 GiB74±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB1.59 GiB128.53 GiB2.60 GiB143±37%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.03 GiB128.36 GiB2.77 GiB138±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB1.00 GiB127.88 GiB3.25 GiB120±37%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB2.92 GiB127.81 GiB3.32 GiB22±22%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB2.92 GiB127.81 GiB3.32 GiB22±22%
MiniMax-M2.7MoEQ4_K_S229B124.65 GiB2.06 GiB127.70 GiB3.43 GiB112±37%
MiniMax-M2.1MoEQ4_K_S229B124.56 GiB2.06 GiB127.61 GiB3.52 GiB112±37%
MiniMax-M2MoEQ4_K_S229B124.56 GiB2.06 GiB127.61 GiB3.52 GiB112±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB1.00 GiB127.57 GiB3.56 GiB120±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB2.06 GiB127.51 GiB3.62 GiB97±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB1.56 GiB127.10 GiB4.03 GiB91±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB1.56 GiB127.10 GiB4.03 GiB91±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB1.56 GiB127.10 GiB4.03 GiB91±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB1.56 GiB127.10 GiB4.03 GiB91±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB1.00 GiB127.00 GiB4.13 GiB120±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB2.06 GiB126.45 GiB4.68 GiB87±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB3.05 GiB126.42 GiB4.71 GiB92±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB3.05 GiB126.42 GiB4.71 GiB92±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB0.56 GiB126.28 GiB4.85 GiB118±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB0.56 GiB126.28 GiB4.85 GiB118±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB0.56 GiB126.28 GiB4.85 GiB118±37%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB0.57 GiB126.03 GiB5.10 GiB125±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB0.57 GiB126.03 GiB5.10 GiB125±37%
Hy3MoEQ3_K_S299B122.26 GiB2.66 GiB125.95 GiB5.18 GiB100±37%
GLM-4.6MoEQ2_K_L357B121.85 GiB3.05 GiB125.94 GiB5.19 GiB93±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB0.92 GiB125.82 GiB5.31 GiB142±37%
Behemoth-X-123B-v2Q8_0123B121.33 GiB2.92 GiB125.41 GiB5.72 GiB22±22%
Mistral-Large-Instruct-2411Q8_0123B121.33 GiB2.92 GiB125.41 GiB5.72 GiB22±22%
QwQ-32BBF1632.8B122.07 GiB2.13 GiB125.29 GiB5.84 GiB22±22%
Ornith-1.0-397BMoEIQ2_M397B123.99 GiB0.25 GiB125.29 GiB5.84 GiB145±37%
Trinity-Large-TrueBaseMoEI1-Q2_K_S399B122.88 GiB0.92 GiB124.82 GiB6.31 GiB143±37%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.20 GiB124.72 GiB6.41 GiB130±37%
grok-2MoEQ3_K_L270B121.28 GiB2.13 GiB124.55 GiB6.58 GiB41±37%
MiniMax-M2.5MoEQ4_K_S229B121.10 GiB2.06 GiB124.15 GiB6.98 GiB114±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_S229B121.10 GiB2.06 GiB124.15 GiB6.98 GiB114±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_M269B119.90 GiB3.05 GiB124.00 GiB7.13 GiB84±37%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
GLM-4.6-Derestricted-v3MoEQ2_K_L357B119.21 GiB3.05 GiB123.31 GiB7.82 GiB94±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB1.79 GiB123.17 GiB7.96 GiB22±22%
Hermes-4-405BIQ2_S406B117.02 GiB4.18 GiB122.48 GiB8.65 GiB23±22%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.20 GiB122.18 GiB8.95 GiB132±37%
Step-3.7-FlashQ4_1201B116.67 GiB3.74 GiB121.44 GiB9.69 GiB23±22%
Laguna-S-2.1MoEQ8_0118B119.91 GiB0.47 GiB121.40 GiB9.73 GiB121±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB0.73 GiB121.38 GiB9.75 GiB115±37%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.25 GiB120.86 GiB10.27 GiB126±37%
step-3.5-flashQ4_1199B115.15 GiB3.74 GiB119.91 GiB11.22 GiB23±22%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.25 GiB119.87 GiB11.26 GiB151±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.19 GiB119.01 GiB12.12 GiB135±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.92 GiB118.45 GiB12.68 GiB149±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB21.25 GiB118.33 GiB12.80 GiB23±22%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.25 GiB117.72 GiB13.41 GiB138±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 16,384 context with q8_0 KV cache, the largest being Qwen3-235B-A22B-Instruct-2507 at Q4_K_S. 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.