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 4K context with q4_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 4K context

largest quantization that fits, per model · 2071 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OYM-Qimi-122B-A10B-K2.6MoEI1-Q4_0125B65.90 GiB0.03 GiB66.95 GiB0.01 GiB72±37%
dots.llm1.instMoEQ3_K_S143B64.78 GiB1.09 GiB66.90 GiB0.06 GiB55±37%
GLM-4.5VMoEI1-Q4_K_M108B65.61 GiB0.20 GiB66.84 GiB0.12 GiB54±37%
Qwen3.5-99BMoEI1-Q5_K_M99.0B65.60 GiB0.03 GiB66.66 GiB0.30 GiB67±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_S235B65.40 GiB0.21 GiB66.64 GiB0.32 GiB54±37%
MiniMax-M2.7MoEUD-IQ2_M229B65.32 GiB0.27 GiB66.58 GiB0.38 GiB69±37%
step-3.5-flashQ2_K199B64.78 GiB0.71 GiB66.52 GiB0.44 GiB12±22%
Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4MoEF1634.7B65.34 GiB0.02 GiB66.37 GiB0.59 GiB72±37%
Ornith-Agents-A1-3.6-35B-A3B-dare_tiesMoEF1634.7B65.34 GiB0.02 GiB66.37 GiB0.59 GiB72±37%
Behemoth-X-123B-v2Q4_K_S123B64.79 GiB0.39 GiB66.33 GiB0.63 GiB12±22%
Mistral-Large-Instruct-2411Q4_K_S123B64.79 GiB0.39 GiB66.33 GiB0.63 GiB12±22%
command-a-plus-05-2026-bf16MoEIQ2_S219B65.10 GiB0.14 GiB66.25 GiB0.71 GiB54±37%
Qwen2.5-Coder-32B-InstructQ8_032.8B64.86 GiB0.28 GiB66.24 GiB0.72 GiB12±22%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_S236B65.07 GiB0.07 GiB66.18 GiB0.78 GiB68±37%
DeepSeek-V2.5MoEIQ2_S236B65.07 GiB0.07 GiB66.18 GiB0.78 GiB68±37%
DeepSeek-Coder-V2-InstructMoEIQ2_S236B65.07 GiB0.07 GiB66.18 GiB0.78 GiB68±37%
Qwen3-VL-235B-A22B-ThinkingMoEUD-IQ1_M236B64.90 GiB0.21 GiB66.14 GiB0.82 GiB54±37%
Mistral-Small-4-119B-2603MoEQ4_K_S119B65.08 GiB0.02 GiB66.14 GiB0.82 GiB72±37%
Qwen3-VL-235B-A22B-InstructMoEUD-IQ1_M236B64.83 GiB0.21 GiB66.07 GiB0.89 GiB54±37%
GLM-4.5-Air-DerestrictedMoEQ4_1110B64.77 GiB0.20 GiB66.00 GiB0.96 GiB54±37%
GLM-4.5-AirMoEQ4_1110B64.77 GiB0.20 GiB66.00 GiB0.96 GiB54±37%
Mistral-Medium-3.5-128BIQ4_XS128B64.39 GiB0.39 GiB65.94 GiB1.02 GiB12±22%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_S123B64.86 GiB0.03 GiB65.91 GiB1.05 GiB73±37%
Qwen3.5-REAP-212B-A17BMoEIQ2_M212B64.76 GiB0.03 GiB65.85 GiB1.11 GiB68±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ1_M310B64.64 GiB0.13 GiB65.82 GiB1.14 GiB72±37%
CalmeRys-78B-Orpo-v0.1Q6_K78.0B64.27 GiB0.38 GiB65.78 GiB1.18 GiB12±22%
calme-2.3-rys-78bQ6_K78.0B64.27 GiB0.38 GiB65.78 GiB1.18 GiB12±22%
Qwen3.6-35B-A3B-uncensored-hereticMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Darwin-35B-A3B-OpusMoEBF1636.0B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Carnice-MoE-35B-A3BMoEF1636.0B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.6-35B-A3B-hereticMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Aurora-Code-1MoEBF1634.7B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
grug-35b-v2MoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
grug-35bMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
WorldSim-Opus-3.6-35B-A3BMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Huihui-Nex-N2-mini-abliteratedMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.6-35B-A3B-AnkoMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
KAT-Coder-V2.5-DevMoEBF1634.7B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Ornith-1.0-35B-uncensored-hereticMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.6-35B-A3B-abliterated-v4MoEBF1634.7B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.5-35B-A3B-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Nex-N2-mini-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Agents-A1MoEF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.6-35B-A3B-java-v1MoEBF1634.7B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Nex-N2-miniMoEBF1635.1B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.5-35B-A3B-BaseMoEBF1636.0B64.61 GiB0.02 GiB65.64 GiB1.32 GiB73±37%
Qwen3.5-REAP-262B-A17BMoEIQ2_XXS262B64.50 GiB0.03 GiB65.59 GiB1.37 GiB74±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ4_1109B64.35 GiB0.21 GiB65.59 GiB1.37 GiB54±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_K_S121B63.83 GiB0.71 GiB65.57 GiB1.39 GiB12±22%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ4_NL124B64.39 GiB0.10 GiB65.48 GiB1.48 GiB65±37%
Laguna-S-2.1MoEQ4_K_S118B64.36 GiB0.09 GiB65.48 GiB1.48 GiB67±37%
Step-3.7-FlashIQ2_M201B63.68 GiB0.71 GiB65.42 GiB1.54 GiB12±22%
MiniMax-M2MoEUD-IQ1_M229B64.01 GiB0.27 GiB65.27 GiB1.69 GiB71±37%
grok-2MoEIQ2_XXS270B63.81 GiB0.28 GiB65.23 GiB1.73 GiB23±37%
MiniMax-M2.1MoEUD-IQ1_M229B63.74 GiB0.27 GiB65.00 GiB1.96 GiB71±37%
MiniMax-M2.5MoEUD-IQ1_M229B63.74 GiB0.27 GiB65.00 GiB1.96 GiB71±37%
Qwen3.5-122B-A10BMoEIQ4_XS125B63.77 GiB0.03 GiB64.83 GiB2.13 GiB74±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ2_S229B63.36 GiB0.27 GiB64.62 GiB2.34 GiB71±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 4,096 context with q4_0 KV cache, the largest being OYM-Qimi-122B-A10B-K2.6 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.