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 64K 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 64K context

largest quantization that fits, per model · 2094 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB6.19 GiB131.08 GiB0.05 GiB21±22%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB6.19 GiB131.07 GiB0.06 GiB21±22%
command-a-plus-05-2026-bf16MoEQ4_1219B129.29 GiB0.68 GiB130.98 GiB0.15 GiB93±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEF3235.1B129.13 GiB0.35 GiB130.49 GiB0.64 GiB123±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ2_XXS402B125.91 GiB3.38 GiB130.31 GiB0.82 GiB123±37%
Qwen3-235B-A22B-Instruct-2507MoEQ4_0235B125.72 GiB3.30 GiB130.06 GiB1.07 GiB82±37%
Qwen3-235B-A22B-Thinking-2507MoEQ4_0235B125.72 GiB3.30 GiB130.06 GiB1.07 GiB82±37%
MiniMax-M2.7MoEQ4_K_S229B124.65 GiB4.36 GiB130.00 GiB1.13 GiB96±37%
MiniMax-M2.1MoEQ4_K_S229B124.56 GiB4.36 GiB129.91 GiB1.22 GiB96±37%
MiniMax-M2MoEQ4_K_S229B124.56 GiB4.36 GiB129.91 GiB1.22 GiB96±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB6.47 GiB129.83 GiB1.30 GiB77±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB6.47 GiB129.83 GiB1.30 GiB77±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ2_S480B124.42 GiB4.36 GiB129.81 GiB1.32 GiB85±37%
DeepSeek-V4-FlashMoEUD-IQ4_NL291B128.43 GiB0.02 GiB129.50 GiB1.63 GiB137±37%
GLM-4.6MoEQ2_K_L357B121.85 GiB6.47 GiB129.36 GiB1.77 GiB77±37%
MiMo-V2.5MoEKV unresolvedQ3_K_S311B125.83 GiB2.11 GiB128.99 GiB2.14 GiB110±37%
Hy3MoEQ3_K_S299B122.26 GiB5.63 GiB128.92 GiB2.21 GiB84±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ4_K_S236B124.51 GiB3.30 GiB128.84 GiB2.29 GiB83±37%
Qwen3-VL-235B-A22B-InstructMoEQ4_K_S236B124.51 GiB3.30 GiB128.84 GiB2.29 GiB83±37%
Qwen3-235B-A22BMoEQ4_K_S235B124.51 GiB3.30 GiB128.84 GiB2.29 GiB83±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q4_K_S235B124.51 GiB3.30 GiB128.84 GiB2.29 GiB83±37%
Qwen3-Coder-REAP-363B-A35BMoEQ2_K_L363B123.36 GiB4.36 GiB128.75 GiB2.38 GiB77±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_M310B125.52 GiB2.11 GiB128.68 GiB2.45 GiB111±37%
Behemoth-X-123B-v2Q8_0123B121.33 GiB6.19 GiB128.67 GiB2.46 GiB21±22%
Mistral-Large-Instruct-2411Q8_0123B121.33 GiB6.19 GiB128.67 GiB2.46 GiB21±22%
DeepSeek-V4-Flash-0731MoEUD-IQ4_NL304B127.28 GiB0.02 GiB128.34 GiB2.79 GiB139±37%
MiniMax-M3MoEUD-IQ2_M427B124.99 GiB2.11 GiB128.12 GiB3.01 GiB111±37%
GLM-4.7-REAP-218B-A32BMoEQ4_K_S218B120.24 GiB6.47 GiB127.75 GiB3.38 GiB66±37%
QwQ-32BBF1632.8B122.07 GiB4.50 GiB127.67 GiB3.46 GiB22±22%
GLM-4.6-REAP-268B-A32BMoEQ3_K_M269B119.90 GiB6.47 GiB127.41 GiB3.72 GiB71±37%
Hermes-4-405BIQ2_S406B117.02 GiB8.86 GiB127.16 GiB3.97 GiB22±22%
grok-2MoEQ3_K_L270B121.28 GiB4.50 GiB126.92 GiB4.21 GiB38±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ4_K_S236B124.68 GiB1.19 GiB126.91 GiB4.22 GiB113±37%
DeepSeek-V2.5MoEQ4_K_S236B124.68 GiB1.19 GiB126.91 GiB4.22 GiB113±37%
DeepSeek-Coder-V2-InstructMoEQ4_K_S236B124.68 GiB1.19 GiB126.91 GiB4.22 GiB113±37%
GLM-4.6-Derestricted-v3MoEQ2_K_L357B119.21 GiB6.47 GiB126.72 GiB4.41 GiB78±37%
granite-34b-code-base-8kF3233.7B125.60 GiB0.00 GiB126.68 GiB4.45 GiB22±22%
DeepSeek-V3-0324MoEIQ1_S685B124.38 GiB1.21 GiB126.67 GiB4.46 GiB119±37%
DeepSeek-R1MoEIQ1_S685B124.38 GiB1.21 GiB126.67 GiB4.46 GiB119±37%
MiniMax-M2.5MoEQ4_K_S229B121.10 GiB4.36 GiB126.45 GiB4.68 GiB98±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ4_K_S229B121.10 GiB4.36 GiB126.45 GiB4.68 GiB98±37%
Trinity-Large-ThinkingMoEIQ2_M399B123.88 GiB1.28 GiB126.18 GiB4.95 GiB138±37%
Ornith-1.0-397BMoEIQ2_M397B123.99 GiB0.53 GiB125.57 GiB5.56 GiB141±37%
Trinity-Large-TrueBaseMoEI1-Q2_K_S399B122.88 GiB1.28 GiB125.18 GiB5.95 GiB139±37%
ERNIE-4.5-300B-A47B-PTQ3_K_S300B120.25 GiB3.80 GiB125.17 GiB5.96 GiB22±22%
Qwen3.5-122B-A10BMoEQ8_0125B123.49 GiB0.42 GiB124.94 GiB6.19 GiB127±37%
Step-3.7-FlashQ4_1201B116.67 GiB7.04 GiB124.74 GiB6.39 GiB22±22%
HunyuanImage-2.1Q8_017.5B122.73 GiB0.00 GiB123.78 GiB7.35 GiB22±22%
step-3.5-flashQ4_1199B115.15 GiB7.04 GiB123.22 GiB7.91 GiB22±22%
Qwen3.5-122B-A10B-hereticMoEQ8_0123B120.95 GiB0.42 GiB122.40 GiB8.73 GiB130±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ8_0124B119.65 GiB1.55 GiB122.19 GiB8.94 GiB109±37%
Laguna-S-2.1MoEQ8_0118B119.91 GiB0.88 GiB121.82 GiB9.31 GiB118±37%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB8.86 GiB121.29 GiB9.84 GiB23±22%
Qwen3.5-REAP-212B-A17BMoEQ4_K_M212B119.56 GiB0.53 GiB121.13 GiB10.00 GiB123±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB17.44 GiB120.31 GiB10.82 GiB55±37%
Qwen3.5-397B-A17BMoEIQ2_S403B118.57 GiB0.53 GiB120.15 GiB10.98 GiB147±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.40 GiB119.21 GiB11.92 GiB133±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB1.28 GiB118.81 GiB12.32 GiB145±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.53 GiB118.00 GiB13.13 GiB134±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.64 GiB117.39 GiB13.74 GiB132±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 65,536 context with q4_0 KV cache, the largest being Devstral-2-123B-Instruct-2512 at Q8_0. 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.