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 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
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◐
Qwen3-Coder-480B-A35B-InstructMoEIQ3_XXS480B176.37 GiB1.09 GiB178.49 GiB0.07 GiB122±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.13 GiB178.12 GiB0.44 GiB169±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.26 GiB178.06 GiB0.50 GiB115±37%
DeepSeek-R1-0528MoEIQ2_S685B176.61 GiB0.30 GiB177.99 GiB0.57 GiB149±37%
DeepSeek-V3.1-TerminusMoEIQ2_S685B176.61 GiB0.30 GiB177.99 GiB0.57 GiB149±37%
DeepSeek-V3.1MoEIQ2_S685B176.61 GiB0.30 GiB177.99 GiB0.57 GiB149±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB0.49 GiB177.94 GiB0.62 GiB173±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB0.49 GiB177.82 GiB0.74 GiB173±37%
MiMo-V2.5MoEKV unresolvedQ4_K_L311B176.24 GiB0.53 GiB177.81 GiB0.75 GiB147±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB0.49 GiB177.75 GiB0.81 GiB173±37%
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB0.13 GiB177.72 GiB0.84 GiB169±37%
Hy3MoEQ4_1299B174.90 GiB1.41 GiB177.34 GiB1.22 GiB130±37%
MiniMax-M2.7MoEUD-Q6_K229B175.15 GiB1.09 GiB177.22 GiB1.34 GiB144±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB1.09 GiB176.99 GiB1.57 GiB144±37%
MiniMax-M2MoEQ6_K229B174.91 GiB1.09 GiB176.99 GiB1.57 GiB144±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB1.09 GiB176.94 GiB1.62 GiB144±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB1.09 GiB176.94 GiB1.62 GiB144±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB0.95 GiB176.64 GiB1.92 GiB26±22%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB0.13 GiB176.47 GiB2.09 GiB170±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_L310B174.72 GiB0.53 GiB176.29 GiB2.27 GiB149±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.15 GiB2.41 GiB137±37%
MiniMax-M3MoEQ3_K_S427B174.39 GiB0.53 GiB175.93 GiB2.63 GiB149±37%
grok-2MoEQ5_K_S270B173.10 GiB1.13 GiB175.37 GiB3.19 GiB49±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_S269B172.64 GiB1.62 GiB175.30 GiB3.26 GiB108±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB0.30 GiB174.83 GiB3.73 GiB151±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB0.30 GiB174.50 GiB4.06 GiB152±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB0.30 GiB174.49 GiB4.07 GiB152±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.13 GiB174.14 GiB4.42 GiB155±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB0.30 GiB173.77 GiB4.79 GiB152±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB0.30 GiB173.48 GiB5.08 GiB153±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB0.30 GiB172.86 GiB5.70 GiB153±37%
Hermes-3-Llama-3.1-405BIQ3_M406B169.26 GiB2.21 GiB172.76 GiB5.80 GiB27±22%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB0.53 GiB172.60 GiB5.96 GiB151±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB0.39 GiB170.76 GiB7.80 GiB154±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB1.62 GiB170.22 GiB8.34 GiB100±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB0.84 GiB167.52 GiB11.04 GiB162±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB0.39 GiB165.48 GiB13.08 GiB158±37%
GLM-5.1MoEIQ1_M754B163.93 GiB0.39 GiB165.37 GiB13.19 GiB158±37%
r1-1776MoEIQ2_XXS671B162.45 GiB0.30 GiB163.83 GiB14.73 GiB161±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB0.30 GiB163.83 GiB14.73 GiB161±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB1.09 GiB163.70 GiB14.86 GiB117±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB1.98 GiB163.21 GiB15.35 GiB28±22%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB0.84 GiB162.67 GiB15.89 GiB199±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB1.62 GiB162.16 GiB16.40 GiB130±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB1.62 GiB162.16 GiB16.40 GiB130±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB1.62 GiB161.54 GiB17.02 GiB130±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB1.62 GiB160.91 GiB17.65 GiB131±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB0.83 GiB157.29 GiB21.27 GiB126±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB0.83 GiB157.29 GiB21.27 GiB126±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB0.83 GiB157.22 GiB21.34 GiB126±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB0.83 GiB157.22 GiB21.34 GiB126±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB0.83 GiB157.22 GiB21.34 GiB126±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB0.83 GiB157.22 GiB21.34 GiB126±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB0.30 GiB157.07 GiB21.49 GiB158±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB0.30 GiB157.07 GiB21.49 GiB158±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB0.30 GiB157.07 GiB21.49 GiB158±37%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.02 GiB154.39 GiB24.17 GiB186±37%
step-3.5-flashQ6_K199B150.81 GiB1.98 GiB153.82 GiB24.74 GiB30±22%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB0.56 GiB151.32 GiB27.24 GiB30±22%
Qwen3-Coder-NextMoEBF1679.7B148.51 GiB0.42 GiB149.92 GiB28.64 GiB193±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 q4_0 KV cache, the largest being Qwen3-Coder-480B-A35B-Instruct at IQ3_XXS. 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.