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 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 32K 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 GiB1.65 GiB130.93 GiB0.20 GiB89±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_S235B128.25 GiB1.65 GiB130.93 GiB0.20 GiB89±37%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB0.40 GiB130.70 GiB0.43 GiB95±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.18 GiB130.31 GiB0.82 GiB125±37%
dots.llm1.instMoEQ6_K143B120.06 GiB8.72 GiB129.80 GiB1.33 GiB71±37%
Solar-Open2-250BMoEIQ4_XS250B126.88 GiB1.69 GiB129.59 GiB1.54 GiB119±37%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.02 GiB129.50 GiB1.63 GiB137±37%
GLM-4.7-REAP-218B-A32BMoEQ4_K_M218B124.65 GiB3.23 GiB128.92 GiB2.21 GiB74±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB1.69 GiB128.63 GiB2.50 GiB142±37%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.02 GiB128.34 GiB2.79 GiB139±37%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB3.09 GiB127.98 GiB3.15 GiB22±22%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB3.09 GiB127.98 GiB3.15 GiB22±22%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB1.05 GiB127.93 GiB3.20 GiB119±37%
MiniMax-M2.7MoEQ4_K_S229B124.65 GiB2.18 GiB127.82 GiB3.31 GiB111±37%
MiniMax-M2.1MoEQ4_K_S229B124.56 GiB2.18 GiB127.73 GiB3.40 GiB111±37%
MiniMax-M2MoEQ4_K_S229B124.56 GiB2.18 GiB127.73 GiB3.40 GiB111±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB2.18 GiB127.63 GiB3.50 GiB96±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB1.05 GiB127.63 GiB3.50 GiB119±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB1.65 GiB127.19 GiB3.94 GiB91±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB1.65 GiB127.19 GiB3.94 GiB91±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB1.65 GiB127.19 GiB3.94 GiB91±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB1.65 GiB127.19 GiB3.94 GiB91±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB1.05 GiB127.06 GiB4.07 GiB120±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
GLM-4.5MoEQ2_K_L358B122.32 GiB3.23 GiB126.60 GiB4.53 GiB91±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB3.23 GiB126.60 GiB4.53 GiB91±37%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB2.18 GiB126.57 GiB4.56 GiB87±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB0.59 GiB126.31 GiB4.82 GiB118±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB0.59 GiB126.31 GiB4.82 GiB118±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB0.59 GiB126.31 GiB4.82 GiB118±37%
GLM-4.6MoEQ2_K_L357B121.85 GiB3.23 GiB126.12 GiB5.01 GiB92±37%
Hy3MoEQ3_K_S299B122.26 GiB2.81 GiB126.11 GiB5.02 GiB99±37%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB0.60 GiB126.07 GiB5.06 GiB125±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB0.60 GiB126.07 GiB5.06 GiB125±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB0.75 GiB125.65 GiB5.48 GiB144±37%
Behemoth-X-123B-v2Q8_0123B121.33 GiB3.09 GiB125.58 GiB5.55 GiB22±22%
Mistral-Large-Instruct-2411Q8_0123B121.33 GiB3.09 GiB125.58 GiB5.55 GiB22±22%
QwQ-32BBF1632.8B122.07 GiB2.25 GiB125.42 GiB5.71 GiB22±22%
Ornith-1.0-397BMoEIQ2_M397B123.99 GiB0.26 GiB125.30 GiB5.83 GiB145±37%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.21 GiB124.73 GiB6.40 GiB130±37%
grok-2MoEQ3_K_L270B121.28 GiB2.25 GiB124.67 GiB6.46 GiB41±37%
Trinity-Large-TrueBaseMoEI1-Q2_K_S399B122.88 GiB0.75 GiB124.65 GiB6.48 GiB145±37%
MiniMax-M2.5MoEQ4_K_S229B121.10 GiB2.18 GiB124.27 GiB6.86 GiB113±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_S229B121.10 GiB2.18 GiB124.27 GiB6.86 GiB113±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_M269B119.90 GiB3.23 GiB124.18 GiB6.95 GiB84±37%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
GLM-4.6-Derestricted-v3MoEQ2_K_L357B119.21 GiB3.23 GiB123.49 GiB7.64 GiB93±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB1.90 GiB123.27 GiB7.86 GiB22±22%
Hermes-4-405BIQ2_S406B117.02 GiB4.43 GiB122.73 GiB8.40 GiB23±22%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.21 GiB122.19 GiB8.94 GiB132±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB0.77 GiB121.42 GiB9.71 GiB115±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB0.46 GiB121.39 GiB9.74 GiB121±37%
Step-3.7-FlashQ4_1201B116.67 GiB3.67 GiB121.37 GiB9.76 GiB23±22%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.26 GiB120.87 GiB10.26 GiB125±37%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.26 GiB119.88 GiB11.25 GiB150±37%
step-3.5-flashQ4_1199B115.15 GiB3.67 GiB119.84 GiB11.29 GiB23±22%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB22.50 GiB119.58 GiB11.55 GiB23±22%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.20 GiB119.02 GiB12.11 GiB135±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.75 GiB118.28 GiB12.85 GiB152±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.26 GiB117.73 GiB13.40 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 32,768 context with q4_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.