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 32K 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 32K context

largest quantization that fits, per model · 2094 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Solar-Open2-250BMoEIQ4_XS250B126.88 GiB3.19 GiB131.09 GiB0.04 GiB107±37%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB0.76 GiB131.05 GiB0.08 GiB93±37%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB5.84 GiB130.73 GiB0.40 GiB21±22%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB5.84 GiB130.73 GiB0.40 GiB21±22%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.33 GiB130.47 GiB0.66 GiB123±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB3.19 GiB130.13 GiB1.00 GiB125±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_0235B125.72 GiB3.12 GiB129.88 GiB1.25 GiB83±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_0235B125.72 GiB3.12 GiB129.88 GiB1.25 GiB83±37%
MiniMax-M2.7MoEQ4_K_S229B124.65 GiB4.12 GiB129.76 GiB1.37 GiB97±37%
MiniMax-M2.1MoEQ4_K_S229B124.56 GiB4.12 GiB129.67 GiB1.46 GiB98±37%
MiniMax-M2MoEQ4_K_S229B124.56 GiB4.12 GiB129.67 GiB1.46 GiB98±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB4.12 GiB129.57 GiB1.56 GiB86±37%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.03 GiB129.51 GiB1.62 GiB137±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB6.11 GiB129.47 GiB1.66 GiB78±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB6.11 GiB129.47 GiB1.66 GiB78±37%
GLM-4.6MoEQ2_K_L357B121.85 GiB6.11 GiB129.00 GiB2.13 GiB79±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB1.99 GiB128.87 GiB2.26 GiB111±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB3.12 GiB128.66 GiB2.47 GiB84±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB3.12 GiB128.66 GiB2.47 GiB84±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB3.12 GiB128.66 GiB2.47 GiB84±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB3.12 GiB128.66 GiB2.47 GiB84±37%
Hy3MoEQ3_K_S299B122.26 GiB5.31 GiB128.61 GiB2.52 GiB86±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB1.99 GiB128.56 GiB2.57 GiB112±37%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB4.12 GiB128.51 GiB2.62 GiB78±37%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.03 GiB128.36 GiB2.77 GiB138±37%
Behemoth-X-123B-v2Q8_0123B121.33 GiB5.84 GiB128.33 GiB2.80 GiB22±22%
Mistral-Large-Instruct-2411Q8_0123B121.33 GiB5.84 GiB128.33 GiB2.80 GiB22±22%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB1.99 GiB128.00 GiB3.13 GiB112±37%
QwQ-32BBF1632.8B122.07 GiB4.25 GiB127.42 GiB3.71 GiB22±22%
GLM-4.7-REAP-218B-A32BMoEQ4_K_S218B120.24 GiB6.11 GiB127.39 GiB3.74 GiB67±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_M269B119.90 GiB6.11 GiB127.05 GiB4.08 GiB73±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB1.12 GiB126.84 GiB4.29 GiB113±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB1.12 GiB126.84 GiB4.29 GiB113±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB1.12 GiB126.84 GiB4.29 GiB113±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
grok-2MoEQ3_K_L270B121.28 GiB4.25 GiB126.67 GiB4.46 GiB39±37%
Hermes-4-405BIQ2_S406B117.02 GiB8.37 GiB126.67 GiB4.46 GiB22±22%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB1.14 GiB126.60 GiB4.53 GiB120±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB1.14 GiB126.60 GiB4.53 GiB120±37%
GLM-4.6-Derestricted-v3MoEQ2_K_L357B119.21 GiB6.11 GiB126.36 GiB4.77 GiB80±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB1.42 GiB126.32 GiB4.81 GiB136±37%
MiniMax-M2.5MoEQ4_K_S229B121.10 GiB4.12 GiB126.21 GiB4.92 GiB99±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_S229B121.10 GiB4.12 GiB126.21 GiB4.92 GiB99±37%
Ornith-1.0-397BMoEIQ2_M397B123.99 GiB0.50 GiB125.54 GiB5.59 GiB142±37%
Trinity-Large-TrueBaseMoEI1-Q2_K_S399B122.88 GiB1.42 GiB125.32 GiB5.81 GiB137±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB3.59 GiB124.96 GiB6.17 GiB22±22%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.40 GiB124.92 GiB6.21 GiB128±37%
Step-3.7-FlashQ4_1201B116.67 GiB6.92 GiB124.63 GiB6.50 GiB22±22%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
step-3.5-flashQ4_1199B115.15 GiB6.92 GiB123.10 GiB8.03 GiB22±22%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.40 GiB122.38 GiB8.75 GiB130±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB1.46 GiB122.11 GiB9.02 GiB110±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB0.87 GiB121.80 GiB9.33 GiB118±37%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.50 GiB121.10 GiB10.03 GiB123±37%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB8.37 GiB120.80 GiB10.33 GiB23±22%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.50 GiB120.12 GiB11.01 GiB147±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB16.47 GiB119.34 GiB11.79 GiB56±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.37 GiB119.19 GiB11.94 GiB133±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB1.42 GiB118.95 GiB12.18 GiB143±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.50 GiB117.97 GiB13.16 GiB135±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 32,768 context with q8_0 KV cache, the largest being Solar-Open2-250B at IQ4_XS. 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.