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 32K context with f16 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 32K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB0.94 GiB178.53 GiB0.03 GiB161±37%
Nex-N2-ProMoEQ3_K_L397B176.46 GiB0.94 GiB178.45 GiB0.11 GiB161±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB0.94 GiB177.27 GiB1.29 GiB162±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB80.00 GiB177.08 GiB1.48 GiB26±22%
DeepSeek-V3.1-TerminusMoEUD-IQ1_S685B173.84 GiB2.14 GiB177.06 GiB1.50 GiB137±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB2.14 GiB176.67 GiB1.89 GiB137±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB2.14 GiB176.35 GiB2.21 GiB137±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB2.14 GiB176.34 GiB2.22 GiB137±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.15 GiB2.41 GiB137±37%
DeepSeek-R1-0528MoEUD-IQ1_S685B172.75 GiB2.14 GiB175.97 GiB2.59 GiB137±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB3.75 GiB175.83 GiB2.73 GiB128±37%
ERNIE-4.5-300B-A47B-PTQ4_K_M300B167.78 GiB6.75 GiB175.66 GiB2.90 GiB26±22%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB2.14 GiB175.61 GiB2.95 GiB138±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB2.14 GiB175.32 GiB3.24 GiB138±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.94 GiB174.95 GiB3.61 GiB148±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB2.14 GiB174.70 GiB3.86 GiB138±37%
Hy3MoEQ4_K_S299B163.41 GiB10.00 GiB174.44 GiB4.12 GiB95±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB13.03 GiB174.26 GiB4.30 GiB26±22%
MiMo-V2.5MoEKV unresolvedQ4_K_S311B169.41 GiB3.75 GiB174.21 GiB4.35 GiB129±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB80.00 GiB174.04 GiB4.52 GiB26±22%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB80.00 GiB174.04 GiB4.52 GiB26±22%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB80.00 GiB174.04 GiB4.52 GiB26±22%
Trinity-Large-PreviewMoEUD-IQ3_XXS399B170.01 GiB2.67 GiB173.70 GiB4.86 GiB155±37%
dots.llm1.instMoEQ8_0143B141.37 GiB31.00 GiB173.40 GiB5.16 GiB56±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB2.74 GiB173.12 GiB5.44 GiB135±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_S310B168.15 GiB3.75 GiB172.95 GiB5.61 GiB129±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB6.00 GiB172.67 GiB5.89 GiB123±37%
Trinity-Large-ThinkingMoEQ3_K_M399B168.90 GiB2.67 GiB172.60 GiB5.96 GiB156±37%
Trinity-Large-TrueBaseMoEQ3_K_M399B168.74 GiB2.67 GiB172.43 GiB6.13 GiB156±37%
MiniMax-M3MoEIQ3_XXS427B167.65 GiB3.75 GiB172.41 GiB6.15 GiB130±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB11.50 GiB172.05 GiB6.51 GiB87±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB11.50 GiB172.05 GiB6.51 GiB87±37%
Qwen3-Coder-480B-A35B-InstructMoEQ2_K_L480B162.87 GiB7.75 GiB171.65 GiB6.91 GiB98±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB11.50 GiB171.43 GiB7.13 GiB88±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB11.50 GiB170.79 GiB7.77 GiB88±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB7.75 GiB170.36 GiB8.20 GiB90±37%
GLM-4.6-REAP-268B-A32BMoEQ4_1269B156.99 GiB11.50 GiB169.53 GiB9.03 GiB81±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB2.74 GiB167.84 GiB10.72 GiB139±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB6.00 GiB167.82 GiB10.74 GiB143±37%
GLM-5.1MoEIQ1_M754B163.93 GiB2.74 GiB167.72 GiB10.84 GiB139±37%
grok-2MoEQ4_1270B157.54 GiB8.00 GiB166.68 GiB11.88 GiB47±37%
MiniMax-M2.7MoEUD-Q5_K_M229B157.23 GiB7.75 GiB165.97 GiB12.59 GiB113±37%
r1-1776MoEIQ2_XXS671B162.45 GiB2.14 GiB165.67 GiB12.89 GiB144±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB2.14 GiB165.67 GiB12.89 GiB144±37%
step-3.5-flashQ6_K199B150.81 GiB13.03 GiB164.87 GiB13.69 GiB28±22%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB5.88 GiB162.33 GiB16.23 GiB102±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB5.88 GiB162.33 GiB16.23 GiB102±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB5.88 GiB162.27 GiB16.29 GiB102±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB5.88 GiB162.27 GiB16.29 GiB102±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB5.88 GiB162.27 GiB16.29 GiB102±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB5.88 GiB162.27 GiB16.29 GiB102±37%
Hermes-3-Llama-3.1-405BIQ3_XXS406B145.14 GiB15.75 GiB162.18 GiB16.38 GiB28±22%
DeepSeek-V3.1MoEUD-TQ1_0685B158.79 GiB2.14 GiB162.01 GiB16.55 GiB147±37%
MiniMax-M2.1MoEQ5_K_M229B151.23 GiB7.75 GiB159.96 GiB18.60 GiB115±37%
MiniMax-M2MoEQ5_K_M229B151.23 GiB7.75 GiB159.96 GiB18.60 GiB115±37%
MiniMax-M2.5MoEQ5_K_M229B151.16 GiB7.75 GiB159.89 GiB18.67 GiB115±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ5_K_M229B151.16 GiB7.75 GiB159.89 GiB18.67 GiB115±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB2.11 GiB158.89 GiB19.67 GiB142±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB2.11 GiB158.89 GiB19.67 GiB142±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB2.11 GiB158.89 GiB19.67 GiB142±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 32,768 context with f16 KV cache, the largest being Ornith-1.0-397B at IQ3_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