NVIDIA · datacenter

B200

B200 has 192 GB of VRAM at 8000 GB/s — about 178.56 GiB usable after driver and compositor overhead. 2104 of 2118 indexed models fit at 128K context with q8_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
text 1808vision language 192image 2audio tts 21audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2104 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.6-Derestricted-v3MoEQ3_K_M357B153.01 GiB24.44 GiB178.48 GiB0.08 GiB62±37%
GLM-4.6MoEIQ3_M357B152.98 GiB24.44 GiB178.46 GiB0.10 GiB62±37%
MiMo-V2.5MoEKV unresolvedQ4_K_S311B169.41 GiB7.97 GiB178.43 GiB0.13 GiB107±37%
DeepSeek-R1-0528MoEUD-IQ1_S685B172.75 GiB4.56 GiB178.38 GiB0.18 GiB122±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB1.99 GiB178.33 GiB0.23 GiB152±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB4.56 GiB178.02 GiB0.54 GiB122±37%
step-3.5-flashQ6_K199B150.81 GiB26.05 GiB177.89 GiB0.67 GiB26±22%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB4.56 GiB177.74 GiB0.82 GiB122±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_S310B168.15 GiB7.97 GiB177.17 GiB1.39 GiB107±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB4.56 GiB177.11 GiB1.45 GiB123±37%
GLM-4.6-REAP-268B-A32BMoEQ4_K_M269B151.37 GiB24.44 GiB176.84 GiB1.72 GiB59±37%
MiniMax-M3MoEIQ3_XXS427B167.65 GiB7.97 GiB176.63 GiB1.93 GiB107±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB5.83 GiB176.20 GiB2.36 GiB117±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.15 GiB2.41 GiB137±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB1.99 GiB176.00 GiB2.56 GiB140±37%
Hermes-3-Llama-3.1-405BQ2_K_L406B140.98 GiB33.47 GiB175.73 GiB2.83 GiB26±22%
grok-2MoEQ4_1270B157.54 GiB17.00 GiB175.68 GiB2.88 GiB40±37%
Trinity-Large-PreviewMoEUD-IQ3_XXS399B170.01 GiB4.40 GiB175.44 GiB3.12 GiB141±37%
MiniMax-M2.7MoEUD-Q5_K_M229B157.23 GiB16.47 GiB174.69 GiB3.87 GiB82±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB12.75 GiB174.57 GiB3.99 GiB105±37%
Trinity-Large-ThinkingMoEQ3_K_M399B168.90 GiB4.40 GiB174.33 GiB4.23 GiB142±37%
Trinity-Large-TrueBaseMoEQ3_K_M399B168.74 GiB4.40 GiB174.17 GiB4.39 GiB142±37%
Hermes-4-405BQ2_K406B139.07 GiB33.47 GiB173.82 GiB4.74 GiB26±22%
ERNIE-4.5-300B-A47B-PTQ4_K_S300B157.93 GiB14.34 GiB173.40 GiB5.16 GiB26±22%
HarmonicHarlequin_v5-20BQ8_033.3B32.97 GiB138.13 GiB172.14 GiB6.42 GiB27±22%
Ornith-1.0-397BMoEQ3_K_M397B169.03 GiB1.99 GiB172.07 GiB6.49 GiB157±37%
Nex-N2-ProMoEQ3_K_M397B169.03 GiB1.99 GiB172.07 GiB6.49 GiB157±37%
Hy3MoEIQ4_XS299B149.73 GiB21.25 GiB172.02 GiB6.54 GiB70±37%
GLM-4.7MoEQ3_K_S358B146.52 GiB24.44 GiB171.99 GiB6.57 GiB63±37%
GLM-4.5MoEQ3_K_S358B146.52 GiB24.44 GiB171.99 GiB6.57 GiB63±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB5.83 GiB170.92 GiB7.64 GiB119±37%
GLM-5.1MoEIQ1_M754B163.93 GiB5.83 GiB170.81 GiB7.75 GiB119±37%
GLM-4.7-REAP-218B-A32BMoEQ5_K_M218B145.23 GiB24.44 GiB170.71 GiB7.85 GiB57±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB12.48 GiB168.94 GiB9.62 GiB81±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB12.48 GiB168.94 GiB9.62 GiB81±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB12.48 GiB168.87 GiB9.69 GiB81±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB12.48 GiB168.87 GiB9.69 GiB81±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB12.48 GiB168.87 GiB9.69 GiB81±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB12.48 GiB168.87 GiB9.69 GiB81±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB65.88 GiB168.75 GiB9.81 GiB34±37%
MiniMax-M2.1MoEQ5_K_M229B151.23 GiB16.47 GiB168.68 GiB9.88 GiB83±37%
MiniMax-M2MoEQ5_K_M229B151.23 GiB16.47 GiB168.68 GiB9.88 GiB83±37%
MiniMax-M2.5MoEQ5_K_M229B151.16 GiB16.47 GiB168.61 GiB9.95 GiB83±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ5_K_M229B151.16 GiB16.47 GiB168.61 GiB9.95 GiB83±37%
DeepSeek-V3.1-TerminusMoEIQ2_XXS685B162.59 GiB4.56 GiB168.23 GiB10.33 GiB127±37%
cogito-671b-v2.1MoEIQ2_XXS671B162.59 GiB4.56 GiB168.23 GiB10.33 GiB127±37%
DeepSeek-V3-0324MoEIQ2_XXS685B162.45 GiB4.56 GiB168.08 GiB10.48 GiB127±37%
r1-1776MoEIQ2_XXS671B162.45 GiB4.56 GiB168.08 GiB10.48 GiB127±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB4.56 GiB168.08 GiB10.48 GiB127±37%
DeepSeek-V3.1MoEUD-TQ1_0685B158.79 GiB4.56 GiB164.43 GiB14.13 GiB130±37%
Step-3.7-FlashUD-Q5_K_M201B136.43 GiB26.05 GiB163.51 GiB15.05 GiB28±22%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_S363B145.81 GiB16.47 GiB163.31 GiB15.25 GiB73±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB4.48 GiB161.26 GiB17.30 GiB126±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB4.48 GiB161.26 GiB17.30 GiB126±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB4.48 GiB161.26 GiB17.30 GiB126±37%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB8.50 GiB159.26 GiB19.30 GiB29±22%
Qwen2.5-72BF1672.7B135.44 GiB21.25 GiB157.82 GiB20.74 GiB29±22%
Kimi-Dev-72BBF1672.7B135.44 GiB21.25 GiB157.82 GiB20.74 GiB29±22%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB21.25 GiB157.82 GiB20.74 GiB29±22%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ1_M480B139.43 GiB16.47 GiB156.94 GiB21.62 GiB80±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?
2104 of 2118 indexed open-weight models fit a B200 at 131,072 context with q8_0 KV cache, the largest being GLM-4.6-Derestricted-v3 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.