NVIDIA · workstation

RTX PRO 4000 Blackwell

RTX PRO 4000 Blackwell has 24 GB of VRAM at 672 GB/s — about 22.32 GiB usable after driver and compositor overhead. 1958 of 2118 indexed models fit at 8K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
24 GB
GDDR7
Bandwidth
672 GB/s
192-bit bus
Tensor FP16
—
dense
TDP
140 W
$1546 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1681vision language 173image 2video 16audio asr 39audio tts 21embedding 26

What fits at 8K context

largest quantization that fits, per model · 1958 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
cogito-v1-preview-qwen-32BI1-Q4_132.8B19.22 GiB2.00 GiB22.32 GiB0.00 GiB18±22%
QwQ-32B-Snowdrop-v0I1-Q4_132.8B19.22 GiB2.00 GiB22.32 GiB0.00 GiB18±22%
DeepSeek-R1-Distill-Qwen-32B-UncensoredI1-Q4_132.8B19.22 GiB2.00 GiB22.32 GiB0.00 GiB18±22%
RoguePlanet-DeepSeek-R1-Qwen-32B-RPI1-Q4_132.8B19.22 GiB2.00 GiB22.32 GiB0.00 GiB18±22%
Llama-3.2-11B-Vision-InstructF1610.7B20.02 GiB1.25 GiB22.31 GiB0.01 GiB18±22%
OpenBuddy-R1-0528-Distill-Qwen3-32B-Preview0-QATQ4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Qwen3-VL-32B-Instruct-ultra-uncensored-hereticI1-Q4_133.4B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Huihui-Qwen3-VL-32B-Instruct-abliteratedI1-Q4_133.4B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
KAT-DevQ4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
ColorGUI-32BI1-Q4_133.4B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Qwen3-VL-32B-InstructQ4_133.4B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Qwen3-VL-32B-ThinkingQ4_133.4B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Qwen3-32B-UncensoredI1-Q4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Qwen3-32BQ4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Qwen3-32B-abliteratedI1-Q4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
DeepSWE-PreviewQ4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
AReaL-boba-2-32BI1-Q4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Assistant_Pepe_32BI1-Q4_132.8B19.22 GiB2.00 GiB22.31 GiB0.01 GiB18±22%
Seed-OSS-36B-Instruct-biprojected-norm-preserving-abliteratedI1-Q4_036.2B19.21 GiB2.00 GiB22.30 GiB0.02 GiB18±22%
Seed-OSS-36B-InstructQ4_036.2B19.21 GiB2.00 GiB22.30 GiB0.02 GiB18±22%
Hermes-4.3-36B-hereticI1-Q4_036.2B19.21 GiB2.00 GiB22.30 GiB0.02 GiB18±22%
Hermes-4.3-36BQ4_036.2B19.21 GiB2.00 GiB22.30 GiB0.02 GiB18±22%
GLM-Z1-Rumination-32B-0414Q4_K_L33.1B19.31 GiB1.91 GiB22.30 GiB0.02 GiB18±22%
Nemotron-Cascade-2-30B-A3BMoEQ4_K_S31.6B20.91 GiB0.41 GiB22.29 GiB0.03 GiB81±37%
Qwen3.6-35B-A3BMoEUD-Q4_K_M36.0B21.11 GiB0.16 GiB22.27 GiB0.05 GiB102±37%
Qwen3.5-35B-A3BMoEQ4_K_L36.0B21.11 GiB0.16 GiB22.27 GiB0.05 GiB102±37%
Salience-1.5-FlashMoEQ5_K_L31.1B20.52 GiB0.75 GiB22.26 GiB0.06 GiB67±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-IQ3_XXS57.3B20.15 GiB1.05 GiB22.23 GiB0.09 GiB62±37%
Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-ThinkingIQ4_XS39.5B20.42 GiB0.75 GiB22.23 GiB0.09 GiB18±22%
Wizard-Vicuna-30B-UncensoredI1-IQ2_XS32.5B8.97 GiB12.19 GiB22.23 GiB0.09 GiB18±22%
archangel_sft-kto_llama30bI1-IQ2_XS32.5B8.97 GiB12.19 GiB22.23 GiB0.09 GiB18±22%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-ThinkingQ6_K23.4B18.64 GiB2.53 GiB22.22 GiB0.10 GiB18±22%
Apertus-70B-Instruct-2509UD-IQ2_XXS70.6B18.54 GiB2.50 GiB22.22 GiB0.10 GiB18±22%
umt5-xxlF325.7B21.17 GiB0.00 GiB22.22 GiB0.10 GiB18±22%
internlm2-math-plus-20bQ8_019.9B19.66 GiB1.50 GiB22.22 GiB0.10 GiB18±22%
Phi-3.5-MoE-instructMoEKV unresolvedQ3_K_L41.9B20.20 GiB1.00 GiB22.21 GiB0.11 GiB47±37%
Yi-34B-200K-DARE-megamerge-v8I1-Q4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
dolphin-2.9.1-yi-1.5-34b-hereticQ4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
dolphin-2.9.1-yi-1.5-34bI1-Q4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
OrionStar-Yi-34B-Chat-LlamaI1-Q4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
Yi-34B-200K-LlamafiedI1-Q4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
Yi-1.5-34BQ4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
Nous-Hermes-2-Yi-34BQ4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
Merged-RP-Stew-V2-34BI1-Q4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
Capybara-Tess-Yi-34B-200KQ4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
Nous-Capybara-limarpv3-34BQ4_K_M34.4B19.24 GiB1.88 GiB22.20 GiB0.12 GiB18±22%
spoomplesmaxx-v2.1-30BI1-Q5_K_M28.9B19.09 GiB2.00 GiB22.20 GiB0.12 GiB18±22%
Huihui-granite-4.1-30b-abliteratedI1-Q5_K_M28.9B19.09 GiB2.00 GiB22.20 GiB0.12 GiB18±22%
granite-4.1-30b-hereticI1-Q5_K_M28.9B19.09 GiB2.00 GiB22.20 GiB0.12 GiB18±22%
granite-4.1-30bQ5_K_M28.9B19.09 GiB2.00 GiB22.20 GiB0.12 GiB18±22%
Qwen3-VL-30B-A3B-ThinkingMoEQ5_K_L31.1B20.43 GiB0.75 GiB22.17 GiB0.15 GiB67±37%
MiroThinker-v1.0-30BMoEQ5_K_L30.5B20.43 GiB0.75 GiB22.17 GiB0.15 GiB67±37%
Qwen3-30B-A3BMoEQ5_K_L30.5B20.43 GiB0.75 GiB22.17 GiB0.15 GiB67±37%
Qwen3-30B-A3B-Instruct-2507MoEQ5_K_L30.5B20.43 GiB0.75 GiB22.17 GiB0.15 GiB67±37%
Qwen3-30B-A3B-Thinking-2507MoEQ5_K_L30.5B20.43 GiB0.75 GiB22.17 GiB0.15 GiB67±37%
Pantheon-Proto-RP-1.8-30B-A3BMoEQ5_K_L30.5B20.43 GiB0.75 GiB22.17 GiB0.15 GiB67±37%
Tongyi-DeepResearch-30B-A3BMoEQ5_K_L30.5B20.43 GiB0.75 GiB22.17 GiB0.15 GiB67±37%
Skywork-R1V3-38BQ5_K_S38.4B21.08 GiB0.00 GiB22.15 GiB0.17 GiB18±22%
command-r-35b-writer-v2I1-IQ2_S35.0B11.03 GiB10.00 GiB22.14 GiB0.18 GiB18±22%
Qwen3.6-27B-uncensored-heretic-v2Q6_K27.4B20.57 GiB0.50 GiB22.13 GiB0.19 GiB18±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.

Measured on this card

third-party benchmarks, aggregated
Workload◍MedianMiddle 50%Runs
Prompt processing4056.81 tok/s3153.61–5016.7116
Text generation126.09 tok/s117.58–132.0912
Benchmarked· n=16

Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from llama.cpp-discussion-15013.

Questions people ask

What AI models can a RTX PRO 4000 Blackwell run?
1958 of 2118 indexed open-weight models fit a RTX PRO 4000 Blackwell at 8,192 context with f16 KV cache, the largest being cogito-v1-preview-qwen-32B at I1-Q4_1. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX PRO 4000 Blackwell actually have?
Its nameplate is 24 GB, but about 22.32 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX PRO 4000 Blackwell fast for local AI?
Its memory bandwidth is 672 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.