NVIDIA · datacenter

B200

B200 has 192 GB of VRAM at 8000 GB/s — about 178.56 GiB usable after driver and compositor overhead. 2108 of 2118 indexed models fit at 128K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
192 GB
HBM3e
Bandwidth
8000 GB/s
8192-bit bus
Tensor FP16
2250 TF
dense
TDP
1000 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
vision language 192text 1812image 2audio tts 21audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB1.05 GiB177.39 GiB1.17 GiB161±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_S685B173.84 GiB2.41 GiB177.33 GiB1.23 GiB135±37%
dots.llm1.instMoEQ8_0143B141.37 GiB34.88 GiB177.28 GiB1.28 GiB51±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB2.41 GiB176.94 GiB1.62 GiB135±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB2.41 GiB176.61 GiB1.95 GiB135±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB2.41 GiB176.60 GiB1.96 GiB135±37%
ERNIE-4.5-300B-A47B-PTQ4_K_M300B167.78 GiB7.59 GiB176.50 GiB2.06 GiB26±22%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB4.22 GiB176.30 GiB2.26 GiB125±37%
DeepSeek-R1-0528MoEUD-IQ1_S685B172.75 GiB2.41 GiB176.24 GiB2.32 GiB136±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.15 GiB2.41 GiB137±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB2.41 GiB175.88 GiB2.68 GiB136±37%
Hy3MoEQ4_K_S299B163.41 GiB11.25 GiB175.69 GiB2.87 GiB91±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB2.41 GiB175.59 GiB2.97 GiB136±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB1.05 GiB175.06 GiB3.50 GiB147±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB13.79 GiB175.02 GiB3.54 GiB26±22%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB2.41 GiB174.97 GiB3.59 GiB136±37%
MiMo-V2.5MoEKV unresolvedQ4_K_S311B169.41 GiB4.22 GiB174.68 GiB3.88 GiB126±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB12.94 GiB173.48 GiB5.08 GiB83±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB12.94 GiB173.48 GiB5.08 GiB83±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB3.08 GiB173.46 GiB5.10 GiB133±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB6.75 GiB173.42 GiB5.14 GiB119±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_S310B168.15 GiB4.22 GiB173.42 GiB5.14 GiB126±37%
Trinity-Large-PreviewMoEUD-IQ3_XXS399B170.01 GiB2.33 GiB173.37 GiB5.19 GiB159±37%
MiniMax-M3MoEIQ3_XXS427B167.65 GiB4.22 GiB172.88 GiB5.68 GiB127±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB12.94 GiB172.86 GiB5.70 GiB84±37%
Qwen3-Coder-480B-A35B-InstructMoEQ2_K_L480B162.87 GiB8.72 GiB172.62 GiB5.94 GiB95±37%
Trinity-Large-ThinkingMoEQ3_K_M399B168.90 GiB2.33 GiB172.26 GiB6.30 GiB159±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB12.94 GiB172.23 GiB6.33 GiB84±37%
Trinity-Large-TrueBaseMoEQ3_K_M399B168.74 GiB2.33 GiB172.10 GiB6.46 GiB159±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB8.72 GiB171.33 GiB7.23 GiB87±37%
Ornith-1.0-397BMoEQ3_K_M397B169.03 GiB1.05 GiB171.13 GiB7.43 GiB166±37%
Nex-N2-ProMoEQ3_K_M397B169.03 GiB1.05 GiB171.13 GiB7.43 GiB166±37%
GLM-4.6-REAP-268B-A32BMoEQ4_1269B156.99 GiB12.94 GiB170.97 GiB7.59 GiB78±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB6.75 GiB168.57 GiB9.99 GiB137±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB3.08 GiB168.18 GiB10.38 GiB136±37%
GLM-5.1MoEIQ1_M754B163.93 GiB3.08 GiB168.07 GiB10.49 GiB136±37%
grok-2MoEQ4_1270B157.54 GiB9.00 GiB167.68 GiB10.88 GiB46±37%
MiniMax-M2.7MoEUD-Q5_K_M229B157.23 GiB8.72 GiB166.94 GiB11.62 GiB108±37%
r1-1776MoEIQ2_XXS671B162.45 GiB2.41 GiB165.94 GiB12.62 GiB142±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB2.41 GiB165.94 GiB12.62 GiB142±37%
step-3.5-flashQ6_K199B150.81 GiB13.79 GiB165.63 GiB12.93 GiB28±22%
Hermes-3-Llama-3.1-405BIQ3_XXS406B145.14 GiB17.72 GiB164.14 GiB14.42 GiB28±22%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB6.61 GiB163.07 GiB15.49 GiB99±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB6.61 GiB163.07 GiB15.49 GiB99±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB6.61 GiB163.00 GiB15.56 GiB99±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB6.61 GiB163.00 GiB15.56 GiB99±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB6.61 GiB163.00 GiB15.56 GiB99±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB6.61 GiB163.00 GiB15.56 GiB99±37%
DeepSeek-V3.1MoEUD-TQ1_0685B158.79 GiB2.41 GiB162.28 GiB16.28 GiB145±37%
MiniMax-M2.1MoEQ5_K_M229B151.23 GiB8.72 GiB160.93 GiB17.63 GiB111±37%
MiniMax-M2MoEQ5_K_M229B151.23 GiB8.72 GiB160.93 GiB17.63 GiB111±37%
MiniMax-M2.5MoEQ5_K_M229B151.16 GiB8.72 GiB160.86 GiB17.70 GiB111±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ5_K_M229B151.16 GiB8.72 GiB160.86 GiB17.70 GiB111±37%
GLM-4.7-REAP-218B-A32BMoEQ5_K_M218B145.23 GiB12.94 GiB159.21 GiB19.35 GiB76±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB2.37 GiB159.15 GiB19.41 GiB140±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB2.37 GiB159.15 GiB19.41 GiB140±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB2.37 GiB159.15 GiB19.41 GiB140±37%
Hermes-4-405BQ2_K406B139.07 GiB17.72 GiB158.07 GiB20.49 GiB29±22%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB4.50 GiB155.26 GiB23.30 GiB29±22%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.02 GiB154.39 GiB24.17 GiB186±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 B200 run?
2108 of 2118 indexed open-weight models fit a B200 at 131,072 context with q4_0 KV cache, the largest being Qwen3.5-397B-A17B at Q3_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a B200 actually have?
Its nameplate is 192 GB, but about 178.56 GiB is available to a model once driver and compositor overhead is accounted for.
Is a B200 fast for local AI?
Its memory bandwidth is 8000 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.