NVIDIA · workstation

RTX PRO 6000 Blackwell Workstation Edition

RTX PRO 6000 Blackwell Workstation Edition has 96 GB of VRAM at 1792 GB/s — about 89.28 GiB usable after driver and compositor overhead. 2082 of 2118 indexed models fit at 32K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
96 GB
GDDR7
Bandwidth
1792 GB/s
512-bit bus
Tensor FP16
504 TF
dense
TDP
600 W
$8565 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
vision language 190text 1788audio tts 21image 2audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2082 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiniMax-M3MoEIQ1_S427B84.31 GiB3.75 GiB89.07 GiB0.21 GiB51±37%
CalmeRys-78B-Orpo-v0.1Q8_078.0B77.16 GiB10.75 GiB89.04 GiB0.24 GiB12±22%
calme-2.3-rys-78bQ8_078.0B77.16 GiB10.75 GiB89.04 GiB0.24 GiB12±22%
Qwen3.5-122B-A10BMoEUD-Q5_K_M125B87.21 GiB0.75 GiB88.99 GiB0.29 GiB66±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ2_XS310B84.00 GiB3.75 GiB88.80 GiB0.48 GiB51±37%
dots.llm1.instMoEQ2_K_L143B56.67 GiB31.00 GiB88.69 GiB0.59 GiB16±37%
Step-3.7-FlashUD-IQ3_S201B74.54 GiB13.03 GiB88.59 GiB0.69 GiB12±22%
c4ai-command-r-plus-08-2024Q6_K104B79.32 GiB8.00 GiB88.50 GiB0.78 GiB12±22%
Qwen3.5-REAP-212B-A17BMoEIQ3_M212B86.49 GiB0.94 GiB88.48 GiB0.80 GiB62±37%
ERNIE-4.5-300B-A47B-PTUD-IQ1_S300B80.54 GiB6.75 GiB88.42 GiB0.86 GiB12±22%
Trinity-Large-ThinkingMoEIQ1_M399B84.63 GiB2.67 GiB88.32 GiB0.96 GiB63±37%
HarmonicHarlequin_v5-20BI1-Q5_K_M33.3B21.98 GiB65.00 GiB88.02 GiB1.26 GiB12±22%
Mixtral-8x22B-Instruct-v0.1MoEQ4_K_M141B79.71 GiB7.00 GiB87.78 GiB1.50 GiB19±37%
Mixtral-8x22B-v0.1MoEQ4_K_M141B79.71 GiB7.00 GiB87.77 GiB1.51 GiB19±37%
Mixtral-8x22B-v0.1MoEQ4_K_M141B79.71 GiB7.00 GiB87.77 GiB1.51 GiB19±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XXS236B84.61 GiB2.11 GiB87.76 GiB1.52 GiB56±37%
DeepSeek-V2.5MoEIQ3_XXS236B84.61 GiB2.11 GiB87.76 GiB1.52 GiB56±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XXS236B84.61 GiB2.11 GiB87.76 GiB1.52 GiB56±37%
GLM-4.7-REAP-218B-A32BMoEQ2_K_L218B75.08 GiB11.50 GiB87.62 GiB1.66 GiB27±37%
Hy3MoEIQ2_XXS299B76.47 GiB10.00 GiB87.51 GiB1.77 GiB33±37%
GLM-4.5MoEIQ1_M358B74.85 GiB11.50 GiB87.39 GiB1.89 GiB30±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q5_K_S124B83.56 GiB2.75 GiB87.30 GiB1.98 GiB51±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q4_K_M139B78.40 GiB7.75 GiB87.14 GiB2.14 GiB35±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q4_K_M139B78.40 GiB7.75 GiB87.14 GiB2.14 GiB35±37%
Ornith-1.0-397BMoEIQ1_M397B85.09 GiB0.94 GiB87.08 GiB2.20 GiB74±37%
GLM-4.7MoEIQ1_M358B74.53 GiB11.50 GiB87.07 GiB2.21 GiB30±37%
step-3.5-flashIQ3_XXS199B72.82 GiB13.03 GiB86.88 GiB2.40 GiB12±22%
Qwen3-VL-235B-A22B-ThinkingMoEQ2_K_L236B79.94 GiB5.88 GiB86.85 GiB2.43 GiB38±37%
Qwen3-VL-235B-A22B-InstructMoEQ2_K_L236B79.94 GiB5.88 GiB86.85 GiB2.43 GiB38±37%
Qwen3-235B-A22BMoEQ2_K_L235B79.94 GiB5.88 GiB86.85 GiB2.43 GiB38±37%
Qwen3-235B-A22B-Instruct-2507MoEQ2_K_L235B79.94 GiB5.88 GiB86.85 GiB2.43 GiB38±37%
Qwen3-235B-A22B-Thinking-2507MoEQ2_K_L235B79.94 GiB5.88 GiB86.85 GiB2.43 GiB38±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q2_K235B79.81 GiB5.88 GiB86.71 GiB2.57 GiB38±37%
GLM-4.6-Derestricted-v3MoEIQ1_M357B74.10 GiB11.50 GiB86.64 GiB2.64 GiB30±37%
GLM-4.6MoEIQ1_M357B74.10 GiB11.50 GiB86.64 GiB2.64 GiB30±37%
MiniMax-M2.7MoEUD-IQ3_S229B77.87 GiB7.75 GiB86.61 GiB2.67 GiB39±37%
MiniMax-M2.1MoEQ2_K_L229B77.72 GiB7.75 GiB86.45 GiB2.83 GiB39±37%
MiniMax-M2.5MoEQ2_K_L229B77.72 GiB7.75 GiB86.45 GiB2.83 GiB39±37%
MiniMax-M2MoEQ2_K_L229B77.72 GiB7.75 GiB86.45 GiB2.83 GiB39±37%
Mistral-Medium-3.5-128BQ4_1128B74.29 GiB11.00 GiB86.45 GiB2.83 GiB12±22%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-Q2_K229B77.58 GiB7.75 GiB86.32 GiB2.96 GiB39±37%
DeepSeek-V4-Flash-0731MoEUD-IQ2_M304B84.68 GiB0.06 GiB85.79 GiB3.49 GiB80±37%
DeepSeek-V4-FlashMoEUD-IQ2_M291B84.68 GiB0.06 GiB85.79 GiB3.49 GiB80±37%
Trinity-Large-TrueBaseMoEI1-IQ1_M399B82.07 GiB2.67 GiB85.77 GiB3.51 GiB64±37%
grok-2MoEUD-TQ1_0270B76.17 GiB8.00 GiB85.31 GiB3.97 GiB19±37%
Devstral-2-123B-Instruct-2512Q4_1125B73.08 GiB11.00 GiB85.24 GiB4.04 GiB12±22%
XORTRON-NXTXPRTXXLI1-Q4_1128B73.08 GiB11.00 GiB85.24 GiB4.04 GiB12±22%
Mistral-Small-4-119B-2603MoEUD-Q5_K_M119B83.04 GiB0.70 GiB84.77 GiB4.51 GiB70±37%
GLM-4.5-Air-DerestrictedMoEQ5_K_M110B77.97 GiB5.75 GiB84.75 GiB4.53 GiB38±37%
GLM-4.5-AirMoEQ5_K_M110B77.97 GiB5.75 GiB84.75 GiB4.53 GiB38±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q4_1121B70.61 GiB13.03 GiB84.67 GiB4.61 GiB12±22%
MiMo-V2.5MoEKV unresolvedIQ2_XXS311B79.83 GiB3.75 GiB84.63 GiB4.65 GiB52±37%
Qwen3.5-397B-A17BMoEIQ1_S403B82.64 GiB0.94 GiB84.63 GiB4.65 GiB76±37%
Hunyuan-A13B-InstructMoEQ8_080.4B79.58 GiB4.00 GiB84.58 GiB4.70 GiB12±22%
Ace-Step1.5BF16160M82.03 GiB1.54 GiB84.56 GiB4.72 GiB12±22%
Laguna-S-2.1MoEUD-Q5_K_M118B81.83 GiB1.64 GiB84.49 GiB4.79 GiB60±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q5_K_M125B82.62 GiB0.75 GiB84.40 GiB4.88 GiB69±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ5_K_M123B82.62 GiB0.75 GiB84.39 GiB4.89 GiB69±37%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q5_K_S117B81.92 GiB1.15 GiB84.06 GiB5.22 GiB67±37%
gpt-oss-120b-abliteratedMoEI1-Q5_K_S117B81.92 GiB1.15 GiB84.06 GiB5.22 GiB67±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.

Measured on this card

third-party benchmarks, aggregated
Workload◍MedianMiddle 50%Runs
Prompt processing14316.57 tok/s9546.25–16645.0038
Text generation267.03 tok/s256.42–275.7724
Benchmarked· n=38

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 6000 Blackwell Workstation Edition run?
2082 of 2118 indexed open-weight models fit a RTX PRO 6000 Blackwell Workstation Edition at 32,768 context with f16 KV cache, the largest being MiniMax-M3 at IQ1_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX PRO 6000 Blackwell Workstation Edition actually have?
Its nameplate is 96 GB, but about 89.28 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX PRO 6000 Blackwell Workstation Edition fast for local AI?
Its memory bandwidth is 1792 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.
RTX PRO 6000 Blackwell Workstation Edition — what AI models can it run locally? — ossmodeldb