NVIDIA · workstation

RTX PRO 5000 Blackwell

RTX PRO 5000 Blackwell has 72 GB of VRAM at 1344 GB/s — about 66.96 GiB usable after driver and compositor overhead. 2070 of 2118 indexed models fit at 16K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
72 GB
GDDR7
Bandwidth
1344 GB/s
384-bit bus
Tensor FP16
295 TF
dense
TDP
300 W
$4569 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
vision language 186text 1780image 2audio asr 39audio tts 21video 16embedding 26

What fits at 16K context

largest quantization that fits, per model · 2070 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ4_K_M109B62.91 GiB3.00 GiB66.93 GiB0.03 GiB41±37%
dots.llm1.instMoEIQ2_S143B50.40 GiB15.50 GiB66.93 GiB0.03 GiB21±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q3_K_M139B62.01 GiB3.88 GiB66.87 GiB0.09 GiB41±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q3_K_M139B62.01 GiB3.88 GiB66.87 GiB0.09 GiB41±37%
Behemoth-X-123B-v2Q3_K_L123B60.12 GiB5.50 GiB66.77 GiB0.19 GiB12±22%
Mistral-Large-Instruct-2411Q3_K_L123B60.12 GiB5.50 GiB66.77 GiB0.19 GiB12±22%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ4_NL124B64.39 GiB1.38 GiB66.76 GiB0.20 GiB55±37%
Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4MoEF1634.7B65.34 GiB0.31 GiB66.66 GiB0.30 GiB69±37%
Ornith-Agents-A1-3.6-35B-A3B-dare_tiesMoEF1634.7B65.34 GiB0.31 GiB66.66 GiB0.30 GiB69±37%
GLM-Z1-Rumination-32B-0414BF1633.1B61.74 GiB3.81 GiB66.64 GiB0.32 GiB12±22%
Llama-3_1-Nemotron-51B-InstructQ3_K_L51.5B25.47 GiB40.00 GiB66.61 GiB0.35 GiB12±22%
GLM-4.6VMoEQ4_1108B62.66 GiB2.88 GiB66.56 GiB0.40 GiB42±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q3_K_L121B58.48 GiB7.03 GiB66.54 GiB0.42 GiB12±22%
HarmonicHarlequin_v5-20BQ8_033.3B32.97 GiB32.50 GiB66.51 GiB0.45 GiB12±22%
Mistral-Small-4-119B-2603MoEQ4_K_S119B65.08 GiB0.35 GiB66.47 GiB0.49 GiB69±37%
GLM-4.5VMoEI1-Q4_1108B62.40 GiB2.88 GiB66.30 GiB0.66 GiB42±37%
Qwen3.5-REAP-212B-A17BMoEIQ2_M212B64.76 GiB0.47 GiB66.28 GiB0.68 GiB64±37%
GLM-4.7-REAP-218B-A32BMoEIQ2_S218B59.49 GiB5.75 GiB66.28 GiB0.68 GiB31±37%
Laguna-S-2.1MoEQ4_K_S118B64.36 GiB0.89 GiB66.28 GiB0.68 GiB60±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_S123B64.86 GiB0.38 GiB66.26 GiB0.70 GiB69±37%
Llama-3_3-Nemotron-Super-49B-v1_5IQ4_XS49.9B25.06 GiB40.00 GiB66.20 GiB0.76 GiB12±22%
Llama-3_3-Nemotron-Super-49B-v1IQ4_XS49.9B25.06 GiB40.00 GiB66.20 GiB0.76 GiB12±22%
Valkyrie-49B-v2.1I1-IQ4_XS49.9B25.03 GiB40.00 GiB66.17 GiB0.79 GiB12±22%
CallerBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Dumpling-Qwen2.5-32BBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
OpenThinker-32BF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
INTELLECT-2BF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
openhands-lm-32b-v0.1BF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
LongWriter-Zero-32BBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
OpenCodeReasoning-Nemotron-32BBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
OpenCodeReasoning-Nemotron-32B-IOIBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Qwen2.5-Coder-32B-Instruct-abliteratedF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
QwQ-32B-ArliAI-RpR-v4BF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
OpenThinker2-32BBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Qwen2.5-Coder-32BF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Qwen2.5-32B-InstructF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
QwQ-32B-PreviewBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Qwen2.5-32b-RP-InkF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
deepseek-r1-qwen-2.5-32B-ablatedBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Rombos-LLM-V2.5-Qwen-32bF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
DeepSeek-R1-Distill-Qwen-32B-abliteratedBF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Qwen2.5-32B-ArliAI-RPMax-v1.3F1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
DeepSeek-R1-Distill-Qwen-32BF1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
Qwen2.5-VL-32B-InstructBF1633.5B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
EVA-Qwen2.5-32B-v0.2F1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
EVA-Qwen2.5-32B-v0.1F1632.8B61.04 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
cogito-v1-preview-qwen-32BBF1632.8B61.03 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
QwQ-32B-Snowdrop-v0BF1632.8B61.03 GiB4.00 GiB66.13 GiB0.83 GiB12±22%
OpenBuddy-R1-0528-Distill-Qwen3-32B-Preview0-QATBF1632.8B61.03 GiB4.00 GiB66.12 GiB0.84 GiB12±22%
KAT-DevBF1632.8B61.03 GiB4.00 GiB66.12 GiB0.84 GiB12±22%
Qwen3-VL-32B-InstructBF1633.4B61.03 GiB4.00 GiB66.12 GiB0.84 GiB12±22%
Qwen3-VL-32B-ThinkingBF1633.4B61.03 GiB4.00 GiB66.12 GiB0.84 GiB12±22%
Qwen3-VL-32B-Instruct-ultra-uncensored-hereticBF1633.4B61.03 GiB4.00 GiB66.12 GiB0.84 GiB12±22%
Qwen3-32BBF1632.8B61.03 GiB4.00 GiB66.12 GiB0.84 GiB12±22%
DeepSWE-PreviewBF1632.8B61.03 GiB4.00 GiB66.12 GiB0.84 GiB12±22%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_XS236B63.99 GiB1.05 GiB66.09 GiB0.87 GiB60±37%
DeepSeek-V2.5MoEIQ2_XS236B63.99 GiB1.05 GiB66.09 GiB0.87 GiB60±37%
DeepSeek-Coder-V2-InstructMoEIQ2_XS236B63.99 GiB1.05 GiB66.09 GiB0.87 GiB60±37%
HuatuoGPT-o1-72BQ6_K72.7B59.93 GiB5.00 GiB66.06 GiB0.90 GiB12±22%
Rombo-LLM-V3.0-Qwen-72bQ6_K72.7B59.93 GiB5.00 GiB66.06 GiB0.90 GiB12±22%
●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 RTX PRO 5000 Blackwell run?
2070 of 2118 indexed open-weight models fit a RTX PRO 5000 Blackwell at 16,384 context with f16 KV cache, the largest being Llama-4-Scout-17B-16E-Instruct at Q4_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX PRO 5000 Blackwell actually have?
Its nameplate is 72 GB, but about 66.96 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX PRO 5000 Blackwell fast for local AI?
Its memory bandwidth is 1344 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.