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. 2071 of 2118 indexed models fit at 8K context with q8_0 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
text 1781vision language 186image 2audio asr 39audio tts 21video 16embedding 26

What fits at 8K context

largest quantization that fits, per model · 2071 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Step-3.5-Flash-REAP-121B-A11BI1-Q4_0121B63.71 GiB2.14 GiB66.88 GiB0.08 GiB12±22%
dots.llm1.instMoEIQ3_XS143B61.71 GiB4.12 GiB66.85 GiB0.11 GiB41±37%
Step-3.7-FlashIQ2_M201B63.68 GiB2.14 GiB66.85 GiB0.11 GiB12±22%
CalmeRys-78B-Orpo-v0.1Q6_K78.0B64.27 GiB1.43 GiB66.83 GiB0.13 GiB12±22%
calme-2.3-rys-78bQ6_K78.0B64.27 GiB1.43 GiB66.83 GiB0.13 GiB12±22%
Qwen3.5-99BMoEI1-Q5_K_M99.0B65.60 GiB0.10 GiB66.73 GiB0.23 GiB66±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ1_M236B64.90 GiB0.78 GiB66.72 GiB0.24 GiB51±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ1_M236B64.83 GiB0.78 GiB66.64 GiB0.32 GiB51±37%
GLM-4.5-Air-DerestrictedMoEQ4_1110B64.77 GiB0.76 GiB66.56 GiB0.40 GiB51±37%
GLM-4.5-AirMoEQ4_1110B64.77 GiB0.76 GiB66.56 GiB0.40 GiB51±37%
command-a-plus-05-2026-bf16MoEIQ2_S219B65.10 GiB0.36 GiB66.47 GiB0.49 GiB53±37%
Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4MoEF1634.7B65.34 GiB0.08 GiB66.43 GiB0.53 GiB71±37%
Ornith-Agents-A1-3.6-35B-A3B-dare_tiesMoEF1634.7B65.34 GiB0.08 GiB66.43 GiB0.53 GiB71±37%
MiniMax-M2.7MoEIQ2_S229B64.38 GiB1.03 GiB66.40 GiB0.56 GiB63±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_S236B65.07 GiB0.28 GiB66.39 GiB0.57 GiB66±37%
DeepSeek-V2.5MoEIQ2_S236B65.07 GiB0.28 GiB66.39 GiB0.57 GiB66±37%
DeepSeek-Coder-V2-InstructMoEIQ2_S236B65.07 GiB0.28 GiB66.39 GiB0.57 GiB66±37%
Mistral-Small-4-119B-2603MoEQ4_K_S119B65.08 GiB0.09 GiB66.21 GiB0.75 GiB72±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ1_M310B64.64 GiB0.50 GiB66.19 GiB0.77 GiB68±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ4_1109B64.35 GiB0.80 GiB66.17 GiB0.79 GiB51±37%
MiniMax-M2MoEUD-IQ1_M229B64.01 GiB1.03 GiB66.03 GiB0.93 GiB63±37%
grok-2MoEIQ2_XXS270B63.81 GiB1.06 GiB66.02 GiB0.94 GiB22±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_S123B64.86 GiB0.10 GiB65.98 GiB0.98 GiB72±37%
Qwen3.5-REAP-212B-A17BMoEIQ2_M212B64.76 GiB0.12 GiB65.94 GiB1.02 GiB67±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_XS235B64.09 GiB0.78 GiB65.91 GiB1.05 GiB51±37%
MiniMax-M2.1MoEUD-IQ1_M229B63.74 GiB1.03 GiB65.75 GiB1.21 GiB63±37%
MiniMax-M2.5MoEUD-IQ1_M229B63.74 GiB1.03 GiB65.75 GiB1.21 GiB63±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ4_NL124B64.39 GiB0.37 GiB65.75 GiB1.21 GiB62±37%
Qwen3.6-35B-A3B-uncensored-hereticMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Darwin-35B-A3B-OpusMoEBF1636.0B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Carnice-MoE-35B-A3BMoEF1636.0B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.6-35B-A3B-hereticMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Aurora-Code-1MoEBF1634.7B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
grug-35b-v2MoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
grug-35bMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
WorldSim-Opus-3.6-35B-A3BMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Huihui-Nex-N2-mini-abliteratedMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.6-35B-A3B-AnkoMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
KAT-Coder-V2.5-DevMoEBF1634.7B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Ornith-1.0-35B-uncensored-hereticMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.6-35B-A3B-abliterated-v4MoEBF1634.7B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.5-35B-A3B-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Nex-N2-mini-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Agents-A1MoEF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.6-35B-A3B-java-v1MoEBF1634.7B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Nex-N2-miniMoEBF1635.1B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.5-35B-A3B-BaseMoEBF1636.0B64.61 GiB0.08 GiB65.70 GiB1.26 GiB72±37%
Qwen3.5-REAP-262B-A17BMoEIQ2_XXS262B64.50 GiB0.12 GiB65.68 GiB1.28 GiB73±37%
Laguna-S-2.1MoEQ4_K_S118B64.36 GiB0.27 GiB65.66 GiB1.30 GiB65±37%
XORTRON-NXTXPRTXXLIQ4_XS128B63.03 GiB1.46 GiB65.65 GiB1.31 GiB12±22%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ2_S229B63.36 GiB1.03 GiB65.38 GiB1.58 GiB64±37%
GLM-4.7-REAP-218B-A32BMoEUD-IQ1_M218B62.63 GiB1.53 GiB65.20 GiB1.76 GiB42±37%
Devstral-2-123B-Instruct-2512IQ4_XS125B62.51 GiB1.46 GiB65.13 GiB1.83 GiB12±22%
Mixtral-8x22B-Instruct-v0.1MoEQ3_K_M141B63.14 GiB0.93 GiB65.13 GiB1.83 GiB22±37%
Mixtral-8x22B-v0.1MoEQ3_K_M141B63.14 GiB0.93 GiB65.13 GiB1.83 GiB22±37%
Mixtral-8x22B-v0.1MoEQ3_K_M141B63.13 GiB0.93 GiB65.12 GiB1.84 GiB22±37%
Mistral-Medium-3.5-128BI1-IQ4_XS128B62.47 GiB1.46 GiB65.09 GiB1.87 GiB12±22%
Qwen3.5-122B-A10BMoEIQ4_XS125B63.77 GiB0.10 GiB64.90 GiB2.06 GiB73±37%
HunyuanImage-2.1Q5_117.5B63.57 GiB0.00 GiB64.62 GiB2.34 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?
2071 of 2118 indexed open-weight models fit a RTX PRO 5000 Blackwell at 8,192 context with q8_0 KV cache, the largest being Step-3.5-Flash-REAP-121B-A11B at I1-Q4_0. 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.