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 32K 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 32K context

largest quantization that fits, per model · 2071 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.7-REAP-218B-A32BMoEUD-IQ1_M218B62.63 GiB3.23 GiB66.91 GiB0.05 GiB36±37%
MiniMax-M2.1MoEUD-IQ1_M229B63.74 GiB2.18 GiB66.90 GiB0.06 GiB54±37%
MiniMax-M2.5MoEUD-IQ1_M229B63.74 GiB2.18 GiB66.90 GiB0.06 GiB54±37%
Qwen3.5-99BMoEI1-Q5_K_M99.0B65.60 GiB0.21 GiB66.84 GiB0.12 GiB65±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_XS235B64.09 GiB1.65 GiB66.78 GiB0.18 GiB46±37%
Devstral-2-123B-Instruct-2512IQ4_XS125B62.51 GiB3.09 GiB66.77 GiB0.19 GiB12±22%
MiMo-V2-FlashMoEKV unresolvedI1-IQ1_M310B64.64 GiB1.05 GiB66.74 GiB0.22 GiB62±37%
Mistral-Medium-3.5-128BI1-IQ4_XS128B62.47 GiB3.09 GiB66.72 GiB0.24 GiB12±22%
XORTRON-NXTXPRTXXLI1-IQ4_XS128B62.47 GiB3.09 GiB66.72 GiB0.24 GiB12±22%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_S236B65.07 GiB0.59 GiB66.70 GiB0.26 GiB63±37%
DeepSeek-V2.5MoEIQ2_S236B65.07 GiB0.59 GiB66.70 GiB0.26 GiB63±37%
DeepSeek-Coder-V2-InstructMoEIQ2_S236B65.07 GiB0.59 GiB66.70 GiB0.26 GiB63±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ2_S229B63.36 GiB2.18 GiB66.53 GiB0.43 GiB55±37%
Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4MoEF1634.7B65.34 GiB0.18 GiB66.52 GiB0.44 GiB70±37%
Ornith-Agents-A1-3.6-35B-A3B-dare_tiesMoEF1634.7B65.34 GiB0.18 GiB66.52 GiB0.44 GiB70±37%
command-a-plus-05-2026-bf16MoEIQ2_S219B65.10 GiB0.40 GiB66.51 GiB0.45 GiB53±37%
dots.llm1.instMoEQ2_K_L143B56.67 GiB8.72 GiB66.41 GiB0.55 GiB30±37%
MiniMax-M2.7MoEIQ2_XS229B63.21 GiB2.18 GiB66.37 GiB0.59 GiB55±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ4_K_L109B63.62 GiB1.69 GiB66.34 GiB0.62 GiB47±37%
Mistral-Small-4-119B-2603MoEQ4_K_S119B65.08 GiB0.20 GiB66.31 GiB0.65 GiB70±37%
GLM-4.5-Air-DerestrictedMoEQ4_K_S110B63.62 GiB1.62 GiB66.26 GiB0.70 GiB47±37%
GLM-4.5-AirMoEQ4_K_S110B63.62 GiB1.62 GiB66.26 GiB0.70 GiB47±37%
Mixtral-8x22B-Instruct-v0.1MoEQ3_K_M141B63.14 GiB1.97 GiB66.17 GiB0.79 GiB21±37%
Mixtral-8x22B-v0.1MoEQ3_K_M141B63.14 GiB1.97 GiB66.17 GiB0.79 GiB21±37%
Mixtral-8x22B-v0.1MoEQ3_K_M141B63.13 GiB1.97 GiB66.16 GiB0.80 GiB21±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ4_NL124B64.39 GiB0.77 GiB66.16 GiB0.80 GiB59±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q4_K_S123B64.86 GiB0.21 GiB66.10 GiB0.86 GiB71±37%
Qwen3.5-REAP-212B-A17BMoEIQ2_M212B64.76 GiB0.26 GiB66.08 GiB0.88 GiB66±37%
Laguna-S-2.1MoEQ4_K_S118B64.36 GiB0.46 GiB65.85 GiB1.11 GiB64±37%
Qwen3.5-REAP-262B-A17BMoEIQ2_XXS262B64.50 GiB0.26 GiB65.82 GiB1.14 GiB71±37%
Qwen3.6-35B-A3B-uncensored-hereticMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Darwin-35B-A3B-OpusMoEBF1636.0B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Carnice-MoE-35B-A3BMoEF1636.0B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Qwen3.6-35B-A3B-hereticMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Aurora-Code-1MoEBF1634.7B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
grug-35b-v2MoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
grug-35bMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
WorldSim-Opus-3.6-35B-A3BMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Qwen3.6-35B-A3B-abliterated-MAXMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Huihui-Nex-N2-mini-abliteratedMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Qwen3.6-35B-A3B-AnkoMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
KAT-Coder-V2.5-DevMoEBF1634.7B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Ornith-1.0-35B-uncensored-hereticMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Qwen3.6-35B-A3B-abliterated-v4MoEBF1634.7B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Qwen3.5-35B-A3B-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Nex-N2-mini-ultra-uncensored-hereticMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Agents-A1MoEF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Qwen3.6-35B-A3B-java-v1MoEBF1634.7B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Nex-N2-miniMoEBF1635.1B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
Qwen3.5-35B-A3B-BaseMoEBF1636.0B64.61 GiB0.18 GiB65.79 GiB1.17 GiB71±37%
GLM-4.6VMoEQ4_1108B62.66 GiB1.62 GiB65.30 GiB1.66 GiB47±37%
Behemoth-X-123B-v2IQ4_XS123B60.94 GiB3.09 GiB65.19 GiB1.77 GiB12±22%
Mistral-Large-Instruct-2411IQ4_XS123B60.94 GiB3.09 GiB65.19 GiB1.77 GiB12±22%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q3_K_M139B62.01 GiB2.18 GiB65.18 GiB1.78 GiB48±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q3_K_M139B62.01 GiB2.18 GiB65.18 GiB1.78 GiB48±37%
GLM-4.5VMoEI1-Q4_1108B62.40 GiB1.62 GiB65.04 GiB1.92 GiB48±37%
Qwen3.5-122B-A10BMoEIQ4_XS125B63.77 GiB0.21 GiB65.01 GiB1.95 GiB72±37%
GLM-Z1-Rumination-32B-0414BF1633.1B61.74 GiB2.14 GiB64.98 GiB1.98 GiB12±22%
Step-3.5-Flash-REAP-121B-A11BI1-IQ4_XS121B60.12 GiB3.67 GiB64.82 GiB2.14 GiB12±22%
GLM-4.6-REAP-268B-A32BMoEUD-TQ1_0269B60.36 GiB3.23 GiB64.64 GiB2.32 GiB41±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 RTX PRO 5000 Blackwell run?
2071 of 2118 indexed open-weight models fit a RTX PRO 5000 Blackwell at 32,768 context with q4_0 KV cache, the largest being GLM-4.7-REAP-218B-A32B at UD-IQ1_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.