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 16K 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 1812vision language 192image 2audio tts 21audio asr 39video 16embedding 26

What fits at 16K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB0.92 GiB178.37 GiB0.19 GiB168±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.49 GiB178.29 GiB0.27 GiB114±37%
MiMo-V2.5MoEKV unresolvedQ4_K_L311B176.24 GiB1.00 GiB178.28 GiB0.28 GiB144±37%
DeepSeek-R1-0528MoEIQ2_S685B176.61 GiB0.57 GiB178.26 GiB0.30 GiB147±37%
DeepSeek-V3.1-TerminusMoEIQ2_S685B176.61 GiB0.57 GiB178.26 GiB0.30 GiB147±37%
DeepSeek-V3.1MoEIQ2_S685B176.61 GiB0.57 GiB178.26 GiB0.30 GiB147±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB0.92 GiB178.25 GiB0.31 GiB168±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.25 GiB178.23 GiB0.33 GiB168±37%
MiniMax-M2.7MoEUD-Q6_K229B175.15 GiB2.06 GiB178.19 GiB0.37 GiB136±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB0.92 GiB178.19 GiB0.37 GiB168±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB2.06 GiB177.96 GiB0.60 GiB137±37%
MiniMax-M2MoEQ6_K229B174.91 GiB2.06 GiB177.96 GiB0.60 GiB137±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB2.06 GiB177.91 GiB0.65 GiB137±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB2.06 GiB177.91 GiB0.65 GiB137±37%
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB0.25 GiB177.84 GiB0.72 GiB168±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB1.79 GiB177.48 GiB1.08 GiB26±22%
MiMo-V2-FlashMoEKV unresolvedQ4_K_L310B174.72 GiB1.00 GiB176.76 GiB1.80 GiB145±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_S269B172.64 GiB3.05 GiB176.74 GiB1.82 GiB102±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB0.25 GiB176.59 GiB1.97 GiB169±37%
MiniMax-M3MoEQ3_K_S427B174.39 GiB1.00 GiB176.40 GiB2.16 GiB145±37%
grok-2MoEQ5_K_S270B173.10 GiB2.13 GiB176.37 GiB2.19 GiB48±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.15 GiB2.41 GiB137±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB0.57 GiB175.10 GiB3.46 GiB149±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB0.57 GiB174.77 GiB3.79 GiB149±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB0.57 GiB174.76 GiB3.80 GiB149±37%
Hermes-3-Llama-3.1-405BIQ3_M406B169.26 GiB4.18 GiB174.73 GiB3.83 GiB26±22%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.25 GiB174.26 GiB4.30 GiB154±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB0.57 GiB174.04 GiB4.52 GiB150±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB0.57 GiB173.75 GiB4.81 GiB150±37%
Hy3MoEQ4_K_L299B169.99 GiB2.66 GiB173.68 GiB4.88 GiB125±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB0.57 GiB173.13 GiB5.43 GiB151±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB1.00 GiB173.07 GiB5.49 GiB147±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB3.05 GiB171.66 GiB6.90 GiB95±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB0.73 GiB171.10 GiB7.46 GiB151±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB1.59 GiB168.27 GiB10.29 GiB155±37%
Qwen3-Coder-480B-A35B-InstructMoEQ2_K_L480B162.87 GiB2.06 GiB165.96 GiB12.60 GiB125±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB0.73 GiB165.82 GiB12.74 GiB155±37%
GLM-5.1MoEIQ1_M754B163.93 GiB0.73 GiB165.71 GiB12.85 GiB155±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB3.74 GiB164.96 GiB13.60 GiB28±22%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB2.06 GiB164.67 GiB13.89 GiB112±37%
r1-1776MoEIQ2_XXS671B162.45 GiB0.57 GiB164.10 GiB14.46 GiB158±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB0.57 GiB164.10 GiB14.46 GiB158±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB3.05 GiB163.60 GiB14.96 GiB121±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB3.05 GiB163.60 GiB14.96 GiB121±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB1.59 GiB163.42 GiB15.14 GiB188±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB3.05 GiB162.98 GiB15.58 GiB122±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB3.05 GiB162.34 GiB16.22 GiB122±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB1.56 GiB158.02 GiB20.54 GiB122±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB1.56 GiB158.02 GiB20.54 GiB122±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB1.56 GiB157.95 GiB20.61 GiB122±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB1.56 GiB157.95 GiB20.61 GiB122±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB1.56 GiB157.95 GiB20.61 GiB122±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB1.56 GiB157.95 GiB20.61 GiB122±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB0.56 GiB157.34 GiB21.22 GiB155±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB0.56 GiB157.34 GiB21.22 GiB155±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB0.56 GiB157.34 GiB21.22 GiB155±37%
step-3.5-flashQ6_K199B150.81 GiB3.74 GiB155.58 GiB22.98 GiB29±22%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.03 GiB154.41 GiB24.15 GiB186±37%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB1.06 GiB151.82 GiB26.74 GiB30±22%
dots.llm1.instMoEQ8_0143B141.37 GiB8.23 GiB150.64 GiB27.92 GiB107±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 16,384 context with q8_0 KV cache, the largest being Trinity-Large-Preview 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.
B200 — what AI models can it run locally? — ossmodeldb