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 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
vision language 191text 1799image 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◐
Ornith-1.0-397BMoEQ2_K397B129.92 GiB0.13 GiB131.10 GiB0.03 GiB140±37%
Nex-N2-ProMoEQ2_K397B129.92 GiB0.13 GiB131.10 GiB0.03 GiB140±37%
Step-3.7-FlashQ5_K_S201B128.01 GiB1.98 GiB131.02 GiB0.11 GiB21±22%
Trinity-Large-TrueBaseMoEQ2_K399B129.44 GiB0.49 GiB130.96 GiB0.17 GiB142±37%
MiniMax-M2.5MoEQ4_K_M229B128.84 GiB1.09 GiB130.92 GiB0.21 GiB117±37%
MiniMax-M2.1MoEI1-Q4_K_M229B128.84 GiB1.09 GiB130.92 GiB0.21 GiB117±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_M229B128.84 GiB1.09 GiB130.92 GiB0.21 GiB117±37%
GLM-4.7-REAP-218B-A32BMoEQ4_1218B128.16 GiB1.62 GiB130.82 GiB0.31 GiB78±37%
Hermes-3-Llama-3.1-405BIQ2_M406B127.28 GiB2.21 GiB130.78 GiB0.35 GiB21±22%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB0.26 GiB130.56 GiB0.57 GiB95±37%
Hy3MoEIQ3_XS299B128.00 GiB1.41 GiB130.44 GiB0.69 GiB105±37%
Yi-1.5-34BF3234.4B128.11 GiB1.05 GiB130.25 GiB0.88 GiB21±22%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.09 GiB130.22 GiB0.91 GiB126±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_K_S235B128.25 GiB0.83 GiB130.10 GiB1.03 GiB93±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_S235B128.25 GiB0.83 GiB130.10 GiB1.03 GiB93±37%
step-3.5-flashQ5_K_S199B126.55 GiB1.98 GiB129.56 GiB1.57 GiB21±22%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.02 GiB129.50 GiB1.63 GiB137±37%
Solar-Open2-250BMoEIQ4_XS250B126.88 GiB0.84 GiB128.75 GiB2.38 GiB128±37%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.02 GiB128.34 GiB2.79 GiB139±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB0.84 GiB127.78 GiB3.35 GiB154±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB0.53 GiB127.41 GiB3.72 GiB124±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB0.53 GiB127.10 GiB4.03 GiB124±37%
MiniMax-M2.7MoEQ4_K_S229B124.65 GiB1.09 GiB126.73 GiB4.40 GiB120±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
MiniMax-M2MoEQ4_K_S229B124.56 GiB1.09 GiB126.64 GiB4.49 GiB120±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB1.09 GiB126.54 GiB4.59 GiB103±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB0.53 GiB126.53 GiB4.60 GiB125±37%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB1.55 GiB126.43 GiB4.70 GiB22±22%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB1.55 GiB126.43 GiB4.70 GiB22±22%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB0.83 GiB126.36 GiB4.77 GiB95±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB0.83 GiB126.36 GiB4.77 GiB95±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB0.83 GiB126.36 GiB4.77 GiB95±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB0.83 GiB126.36 GiB4.77 GiB95±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB0.30 GiB126.02 GiB5.11 GiB120±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB0.30 GiB126.02 GiB5.11 GiB120±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB0.30 GiB126.02 GiB5.11 GiB120±37%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB0.30 GiB125.76 GiB5.37 GiB128±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB0.30 GiB125.76 GiB5.37 GiB128±37%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB1.09 GiB125.48 GiB5.65 GiB92±37%
dots.llm1.instMoEQ6_K143B120.06 GiB4.36 GiB125.44 GiB5.69 GiB89±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB0.49 GiB125.39 GiB5.74 GiB148±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB1.62 GiB124.98 GiB6.15 GiB101±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB1.62 GiB124.98 GiB6.15 GiB101±37%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.11 GiB124.62 GiB6.51 GiB131±37%
GLM-4.6MoEQ2_K_L357B121.85 GiB1.62 GiB124.50 GiB6.63 GiB101±37%
QwQ-32BBF1632.8B122.07 GiB1.13 GiB124.29 GiB6.84 GiB22±22%
Behemoth-X-123B-v2Q8_0123B121.33 GiB1.55 GiB124.03 GiB7.10 GiB22±22%
Mistral-Large-Instruct-2411Q8_0123B121.33 GiB1.55 GiB124.03 GiB7.10 GiB22±22%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
grok-2MoEQ3_K_L270B121.28 GiB1.13 GiB123.55 GiB7.58 GiB42±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_M269B119.90 GiB1.62 GiB122.56 GiB8.57 GiB91±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB0.95 GiB122.32 GiB8.81 GiB23±22%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.11 GiB122.09 GiB9.04 GiB133±37%
GLM-4.6-Derestricted-v3MoEQ2_K_L357B119.21 GiB1.62 GiB121.87 GiB9.26 GiB103±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB0.25 GiB121.18 GiB9.95 GiB123±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB0.39 GiB121.03 GiB10.10 GiB118±37%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.13 GiB120.74 GiB10.39 GiB127±37%
Hermes-4-405BIQ2_S406B117.02 GiB2.21 GiB120.52 GiB10.61 GiB23±22%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.13 GiB119.75 GiB11.38 GiB152±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.10 GiB118.92 GiB12.21 GiB136±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 q4_0 KV cache, the largest being Ornith-1.0-397B 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.