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. 2046 of 2118 indexed models fit at 128K 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 1757vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2046 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ3_XS124B60.11 GiB5.84 GiB66.95 GiB0.01 GiB36±37%
Qwen3.6-35B-A3B-uncensored-hereticMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Darwin-35B-A3B-OpusMoEBF1636.0B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Carnice-MoE-35B-A3BMoEF1636.0B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Qwen3.6-35B-A3B-hereticMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Aurora-Code-1MoEBF1634.7B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
grug-35b-v2MoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
grug-35bMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
WorldSim-Opus-3.6-35B-A3BMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Huihui-Nex-N2-mini-abliteratedMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Qwen3.6-35B-A3B-AnkoMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
KAT-Coder-V2.5-DevMoEBF1634.7B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Ornith-1.0-35B-uncensored-hereticMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Qwen3.6-35B-A3B-abliterated-v4MoEBF1634.7B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Qwen3.5-35B-A3B-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Nex-N2-mini-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Agents-A1MoEF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Qwen3.6-35B-A3B-java-v1MoEBF1634.7B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Nex-N2-miniMoEBF1635.1B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
Qwen3.5-35B-A3B-BaseMoEBF1636.0B64.61 GiB1.33 GiB66.94 GiB0.02 GiB60±37%
GLM-4.5VMoEI1-Q3_K_L108B53.68 GiB12.22 GiB66.93 GiB0.03 GiB23±37%
Laguna-S-2.1MoEQ4_0118B62.44 GiB3.26 GiB66.72 GiB0.24 GiB46±37%
Qwen3-14B-GPT-5.2-High-Reasoning-DistillBF1614.8B55.03 GiB10.63 GiB66.71 GiB0.25 GiB12±22%
Behemoth-X-123B-v2Q2_K123B42.09 GiB23.38 GiB66.62 GiB0.34 GiB12±22%
Mistral-Large-Instruct-2411Q2_K123B42.09 GiB23.38 GiB66.62 GiB0.34 GiB12±22%
CodeLlama-70b-Instruct-hfI1-Q5_K_S69.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
CodeLlama-70b-Python-hfI1-Q5_K_S69.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
Nous-Hermes-Llama2-70bI1-Q5_K_S69.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
Midnight-Miqu-70B-v1.5I1-Q5_K_S69.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
KafkaLM-70B-German-V0.1Q5_069.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
llama2_70b_chat_uncensoredQ5_069.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
Xwin-LM-70b-V0.1Q5_069.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
Llama-2-70b-chat-hfQ5_069.0B44.20 GiB21.25 GiB66.57 GiB0.39 GiB12±22%
Rombo-LLM-V3.0-Qwen-72bI1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Qwen2.5-72B-Instruct-abliteratedI1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Qwen2.5-72B-Instruct-abliterated-v2I1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
HuatuoGPT-o1-72BQ4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
MiroThinker-v1.0-72BI1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
EVA-Qwen2.5-72B-v0.2Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Qwen2.5-Math-72B-InstructQ4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Qwen2.5-72B-InstructQ4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Malaysian-Qwen2.5-72B-InstructI1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Qwen2.5-72BI1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
magnum-v4-72bI1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
KAT-Dev-72B-ExpQ4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Homer-v1.0-Qwen2.5-72BQ4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Chuluun-Qwen2.5-72B-v0.01Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Qwen2.5-VL-72B-InstructQ4_K_M73.4B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
Chronos-Platinum-72BQ4_K_M72.7B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
UI-TARS-72B-DPOQ4_K_M73.4B44.16 GiB21.25 GiB66.54 GiB0.42 GiB12±22%
GLM-4.5-AirMoEQ3_K_M110B53.28 GiB12.22 GiB66.52 GiB0.44 GiB24±37%
Qwen3-72B-SynthesisQ4_K_M72.7B44.12 GiB21.25 GiB66.50 GiB0.46 GiB12±22%
Qwen3.5-122B-A10BMoEIQ4_XS125B63.77 GiB1.59 GiB66.39 GiB0.57 GiB59±37%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPIQ4_XS27.8B61.06 GiB4.25 GiB66.37 GiB0.59 GiB12±22%
GLM-4.5-Air-REAP-82B-A12BMoEQ4_K_L81.9B53.11 GiB12.22 GiB66.36 GiB0.60 GiB22±37%
Phi-3-medium-128k-instructF3214.0B52.01 GiB13.28 GiB66.35 GiB0.61 GiB12±22%
Phi-3-medium-4k-instructF3214.0B52.01 GiB13.28 GiB66.35 GiB0.61 GiB12±22%
UncensoredLM-DeepSeek-R1-Distill-Qwen-14BF3214.2B52.98 GiB12.22 GiB66.24 GiB0.72 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?
2046 of 2118 indexed open-weight models fit a RTX PRO 5000 Blackwell at 131,072 context with q8_0 KV cache, the largest being NVIDIA-Nemotron-3-Super-120B-A12B-BF16 at IQ3_XS. 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.