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 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
vision language 191text 1799image 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◐
Ornith-1.0-397BMoEQ2_K397B129.92 GiB0.12 GiB131.09 GiB0.04 GiB140±37%
Nex-N2-ProMoEQ2_K397B129.92 GiB0.12 GiB131.09 GiB0.04 GiB140±37%
MiniMax-M2.1MoEQ4_K_M229B129.07 GiB1.03 GiB131.08 GiB0.05 GiB117±37%
MiniMax-M2MoEQ4_K_M229B129.07 GiB1.03 GiB131.08 GiB0.05 GiB117±37%
MiniMax-M2.5MoEQ4_K_M229B128.84 GiB1.03 GiB130.86 GiB0.27 GiB117±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_M229B128.84 GiB1.03 GiB130.86 GiB0.27 GiB117±37%
GLM-4.7-REAP-218B-A32BMoEQ4_1218B128.16 GiB1.53 GiB130.73 GiB0.40 GiB78±37%
Hermes-3-Llama-3.1-405BIQ2_M406B127.28 GiB2.09 GiB130.65 GiB0.48 GiB21±22%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB0.36 GiB130.65 GiB0.48 GiB95±37%
Hy3MoEIQ3_XS299B128.00 GiB1.33 GiB130.36 GiB0.77 GiB106±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.08 GiB130.22 GiB0.91 GiB126±37%
Yi-1.5-34BF3234.4B128.11 GiB1.00 GiB130.19 GiB0.94 GiB21±22%
Qwen3-235B-A22B-Instruct-2507MoEQ4_K_S235B128.25 GiB0.78 GiB130.06 GiB1.07 GiB93±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_K_S235B128.25 GiB0.78 GiB130.06 GiB1.07 GiB93±37%
step-3.5-flashQ5_K_S199B126.55 GiB2.14 GiB129.72 GiB1.41 GiB21±22%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.03 GiB129.51 GiB1.62 GiB137±37%
Solar-Open2-250BMoEIQ4_XS250B126.88 GiB0.80 GiB128.70 GiB2.43 GiB128±37%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.03 GiB128.36 GiB2.77 GiB138±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB0.80 GiB127.73 GiB3.40 GiB155±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB0.50 GiB127.38 GiB3.75 GiB124±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB0.50 GiB127.07 GiB4.06 GiB125±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
MiniMax-M2.7MoEQ4_K_S229B124.65 GiB1.03 GiB126.67 GiB4.46 GiB120±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB0.50 GiB126.50 GiB4.63 GiB125±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB1.03 GiB126.48 GiB4.65 GiB103±37%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB1.46 GiB126.35 GiB4.78 GiB22±22%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB1.46 GiB126.35 GiB4.78 GiB22±22%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB0.78 GiB126.32 GiB4.81 GiB96±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB0.78 GiB126.32 GiB4.81 GiB96±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB0.78 GiB126.32 GiB4.81 GiB96±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB0.78 GiB126.32 GiB4.81 GiB96±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB0.28 GiB126.00 GiB5.13 GiB121±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB0.28 GiB126.00 GiB5.13 GiB121±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB0.28 GiB126.00 GiB5.13 GiB121±37%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB0.28 GiB125.75 GiB5.38 GiB128±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB0.28 GiB125.75 GiB5.38 GiB128±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB0.67 GiB125.57 GiB5.56 GiB145±37%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB1.03 GiB125.42 GiB5.71 GiB92±37%
dots.llm1.instMoEQ6_K143B120.06 GiB4.12 GiB125.20 GiB5.93 GiB90±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB1.53 GiB124.89 GiB6.24 GiB102±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB1.53 GiB124.89 GiB6.24 GiB102±37%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.10 GiB124.62 GiB6.51 GiB131±37%
Trinity-Large-TrueBaseMoEI1-Q2_K_S399B122.88 GiB0.67 GiB124.57 GiB6.56 GiB146±37%
GLM-4.6MoEQ2_K_L357B121.85 GiB1.53 GiB124.41 GiB6.72 GiB102±37%
QwQ-32BBF1632.8B122.07 GiB1.06 GiB124.23 GiB6.90 GiB22±22%
Behemoth-X-123B-v2Q8_0123B121.33 GiB1.46 GiB123.95 GiB7.18 GiB22±22%
Mistral-Large-Instruct-2411Q8_0123B121.33 GiB1.46 GiB123.95 GiB7.18 GiB22±22%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
grok-2MoEQ3_K_L270B121.28 GiB1.06 GiB123.48 GiB7.65 GiB42±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_M269B119.90 GiB1.53 GiB122.47 GiB8.66 GiB92±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB0.90 GiB122.27 GiB8.86 GiB23±22%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.10 GiB122.08 GiB9.05 GiB133±37%
GLM-4.6-Derestricted-v3MoEQ2_K_L357B119.21 GiB1.53 GiB121.78 GiB9.35 GiB104±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB0.27 GiB121.21 GiB9.92 GiB123±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB0.37 GiB121.01 GiB10.12 GiB119±37%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.12 GiB120.73 GiB10.40 GiB127±37%
Hermes-4-405BIQ2_S406B117.02 GiB2.09 GiB120.39 GiB10.74 GiB23±22%
Step-3.7-FlashQ4_1201B116.67 GiB2.14 GiB119.84 GiB11.29 GiB23±22%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.12 GiB119.75 GiB11.38 GiB152±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.09 GiB118.91 GiB12.22 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 q8_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.