NVIDIA · workstation

RTX PRO 2000 Blackwell

RTX PRO 2000 Blackwell has 16 GB of VRAM at 288 GB/s — about 14.88 GiB usable after driver and compositor overhead. 1687 of 2118 indexed models fit at 32K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
16 GB
GDDR7
Bandwidth
288 GB/s
128-bit bus
Tensor FP16
—
dense
TDP
70 W
$722 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1434vision language 151audio asr 39video 15embedding 26audio tts 21image 1

What fits at 32K context

largest quantization that fits, per model · 1687 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Gemma-4-31B-Isometry-RPI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Gemma-4-Dark-Gemistry-31BI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Prosopon-31BI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Gemma-4-Novelist-Eclipse-31BI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Giftige-Blume-31B-v1-StyleSwapI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
G4-MeroMero-31B-StyleSwapI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Gemma-4-31B-StyleTune-heretic-araI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Pantheon-Reasoning-31B-1.1I1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Gemma-4-31B-StyleTuneI1-IQ1_M32.7B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Barcenas-StyleTune-31B-FableI1-IQ1_M32.1B7.63 GiB6.17 GiB14.88 GiB0.00 GiB12±22%
Qwen3.6-27B-A3B-CoderMoEI1-IQ4_XS26.7B13.25 GiB0.63 GiB14.87 GiB0.01 GiB44±37%
granite-20b-code-instruct-8kQ5_K_M20.1B13.79 GiB0.00 GiB14.87 GiB0.01 GiB12±22%
granite-20b-code-base-8kI1-Q5_K_M20.1B13.79 GiB0.00 GiB14.87 GiB0.01 GiB12±22%
granite-34b-code-base-8kI1-IQ3_S33.7B13.79 GiB0.00 GiB14.87 GiB0.01 GiB12±22%
gemma-4-A4B-98e-v7-coder-itMoEQ4_K_M20.5B12.33 GiB1.54 GiB14.86 GiB0.02 GiB12±22%
gemma-4-A4B-98e-v7-coderx-itMoEQ4_K_M20.5B12.33 GiB1.54 GiB14.86 GiB0.02 GiB12±22%
Skyfall-31B-v4.2-hereticI1-IQ1_M31.4B6.99 GiB6.75 GiB14.86 GiB0.02 GiB12±22%
Skyfall-31B-v4.2I1-IQ1_M31.4B6.99 GiB6.75 GiB14.86 GiB0.02 GiB12±22%
Skywork-R1V3-38BIQ3_M38.4B13.79 GiB0.00 GiB14.86 GiB0.02 GiB12±22%
Ling-mini-2.0MoEQ6_K_L16.3B12.61 GiB1.25 GiB14.85 GiB0.03 GiB39±37%
internlm2-math-plus-20bI1-IQ3_XS19.9B7.79 GiB6.00 GiB14.85 GiB0.03 GiB12±22%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEI1-Q4_023.0B12.18 GiB1.65 GiB14.85 GiB0.03 GiB28±37%
UncensoredLM-DeepSeek-R1-Distill-Qwen-14BQ4_K_M14.2B8.05 GiB5.75 GiB14.84 GiB0.04 GiB12±22%
spoomplesmaxx-v2.1-30BI1-IQ1_S28.9B5.73 GiB8.00 GiB14.84 GiB0.04 GiB12±22%
Huihui-granite-4.1-30b-abliteratedI1-IQ1_S28.9B5.73 GiB8.00 GiB14.84 GiB0.04 GiB12±22%
granite-4.1-30b-hereticI1-IQ1_S28.9B5.73 GiB8.00 GiB14.84 GiB0.04 GiB12±22%
Llama3.2-24B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q4_K_M18.0B10.33 GiB3.50 GiB14.84 GiB0.04 GiB18±37%
Apriel-1.6-15b-ThinkerI1-Q4_K_S14.9B7.78 GiB6.00 GiB14.83 GiB0.05 GiB12±22%
NuExtract-1.5Q3_K_M3.8B1.82 GiB12.00 GiB14.83 GiB0.05 GiB12±22%
Phi-3.5-mini-instructQ3_K_M3.8B1.82 GiB12.00 GiB14.83 GiB0.05 GiB12±22%
Phi-3.5-mini-instruct_UncensoredQ3_K_M3.8B1.82 GiB12.00 GiB14.83 GiB0.05 GiB12±22%
Phi-3-mini-128k-instructQ3_K_M3.8B1.82 GiB12.00 GiB14.83 GiB0.05 GiB12±22%
Phi-3-mini-4k-instructQ3_K_M3.8B1.82 GiB12.00 GiB14.83 GiB0.05 GiB12±22%
octo-netQ3_K3.8B1.82 GiB12.00 GiB14.83 GiB0.05 GiB12±22%
gemma-4-26B-A4B-itMoEQ3_K_L26.5B12.29 GiB1.54 GiB14.82 GiB0.06 GiB12±22%
GRM-2.6-Plus-0628IQ3_XXS27.8B11.76 GiB2.00 GiB14.82 GiB0.06 GiB12±22%
ThinkingCap-Qwen3.6-27BIQ3_XXS27.4B11.76 GiB2.00 GiB14.82 GiB0.06 GiB12±22%
Tess-4-27BIQ3_XXS27.8B11.76 GiB2.00 GiB14.82 GiB0.06 GiB12±22%
Qwen3-16B-A3BMoEQ5_K_L16.0B10.83 GiB3.00 GiB14.82 GiB0.06 GiB19±37%
NousCoder-14BQ4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
spoomplesmaxx-mini-14BI1-Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
vanilla-cn-roleplay-0.2I1-Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Claria-14bI1-Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
NTX-2.1-ProI1-Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Qwen3-14B-UncensoredI1-Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Qwen3-14BQ4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
FrogMini-14B-2510I1-Q4_1—8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Qwen3-14B-abliteratedQ4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Josiefied-Qwen3-14B-abliterated-v3Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Hermes-4-14BQ4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Slava-Qwen3-14B-SerbianI1-Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
Huihui-Qwen3-14B-abliterated-v2I1-Q4_114.8B8.74 GiB5.00 GiB14.80 GiB0.08 GiB12±22%
NVIDIA-Nemotron-Nano-9B-v2Q5_K_L8.9B6.76 GiB7.00 GiB14.80 GiB0.08 GiB12±22%
GLM-4.7-Flash-DerestrictedMoEI1-IQ3_S31.2B12.14 GiB1.65 GiB14.80 GiB0.08 GiB30±37%
Huihui-GLM-4.7-Flash-abliteratedMoEI1-IQ3_S31.2B12.14 GiB1.65 GiB14.80 GiB0.08 GiB30±37%
GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-DistillMoEQ3_K_S31.2B12.14 GiB1.65 GiB14.80 GiB0.08 GiB30±37%
Pantheon-Reasoning-27BI1-IQ3_S27.8B11.74 GiB2.00 GiB14.80 GiB0.08 GiB12±22%
Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-PreservedI1-IQ3_S27.4B11.74 GiB2.00 GiB14.80 GiB0.08 GiB12±22%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPI1-IQ3_S27.8B11.74 GiB2.00 GiB14.80 GiB0.08 GiB12±22%
Qwen3.6-27B-Fable-5-ExperimentalI1-IQ3_S27.8B11.74 GiB2.00 GiB14.80 GiB0.08 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 2000 Blackwell run?
1687 of 2118 indexed open-weight models fit a RTX PRO 2000 Blackwell at 32,768 context with f16 KV cache, the largest being Gemma-4-31B-Isometry-RP at I1-IQ1_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX PRO 2000 Blackwell actually have?
Its nameplate is 16 GB, but about 14.88 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX PRO 2000 Blackwell fast for local AI?
Its memory bandwidth is 288 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.