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. 2087 of 2118 indexed models fit at 32K context with q4_0 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
text 1793vision language 190audio tts 21image 2audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2087 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
xLAM-8x7b-rMoEBF1646.7B86.99 GiB1.13 GiB89.16 GiB0.12 GiB22±37%
Mistral-Medium-3.5-128BQ5_K_M128B84.85 GiB3.09 GiB89.10 GiB0.18 GiB12±22%
command-a-plus-05-2026-bf16MoEIQ3_XXS219B87.48 GiB0.40 GiB88.88 GiB0.40 GiB53±37%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q5_K_M117B87.49 GiB0.32 GiB88.80 GiB0.48 GiB69±37%
gpt-oss-120b-abliteratedMoEI1-Q5_K_M117B87.49 GiB0.32 GiB88.80 GiB0.48 GiB69±37%
MiMo-V2.5MoEKV unresolvedIQ2_XS311B86.68 GiB1.05 GiB88.78 GiB0.50 GiB64±37%
step-3.5-flashQ3_K_M199B84.04 GiB3.67 GiB88.73 GiB0.55 GiB12±22%
Hy3MoEIQ2_XS299B84.81 GiB2.81 GiB88.65 GiB0.63 GiB51±37%
Qwen3.5-122B-A10BMoEUD-Q5_K_M125B87.21 GiB0.21 GiB88.45 GiB0.83 GiB71±37%
Step-3.7-FlashUD-Q3_K_M201B83.13 GiB3.67 GiB87.83 GiB1.45 GiB12±22%
Qwen3.5-REAP-212B-A17BMoEIQ3_M212B86.49 GiB0.26 GiB87.81 GiB1.47 GiB67±37%
GLM-4.7-REAP-218B-A32BMoEUD-IQ3_XXS218B83.34 GiB3.23 GiB87.62 GiB1.66 GiB39±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ2_S310B85.29 GiB1.05 GiB87.39 GiB1.89 GiB65±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ2_XXS269B83.03 GiB3.23 GiB87.30 GiB1.98 GiB43±37%
GLM-4.5MoEIQ2_XXS358B83.01 GiB3.23 GiB87.28 GiB2.00 GiB47±37%
Gemma-4-Dark-Gemistry-31BQ6_K32.7B84.44 GiB1.74 GiB87.26 GiB2.02 GiB12±22%
dots.llm1.instMoEQ4_0143B77.44 GiB8.72 GiB87.18 GiB2.10 GiB34±37%
MiniMax-M2.1MoEIQ3_XXS229B83.91 GiB2.18 GiB87.08 GiB2.20 GiB59±37%
MiniMax-M2MoEIQ3_XXS229B83.91 GiB2.18 GiB87.08 GiB2.20 GiB59±37%
GLM-4.7MoEIQ2_XXS358B82.69 GiB3.23 GiB86.96 GiB2.32 GiB47±37%
Hermes-4-405BUD-IQ1_S406B81.23 GiB4.43 GiB86.94 GiB2.34 GiB12±22%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_XXS235B84.17 GiB1.65 GiB86.85 GiB2.43 GiB49±37%
GLM-4.5-Air-REAP-82B-A12BMoEQ8_081.9B84.17 GiB1.62 GiB86.81 GiB2.47 GiB43±37%
GLM-4.6-Derestricted-v3MoEIQ2_XXS357B82.26 GiB3.23 GiB86.53 GiB2.75 GiB48±37%
GLM-4.6MoEIQ2_XXS357B82.26 GiB3.23 GiB86.53 GiB2.75 GiB48±37%
Devstral-2-123B-Instruct-2512Q5_K_M125B82.25 GiB3.09 GiB86.50 GiB2.78 GiB12±22%
XORTRON-NXTXPRTXXLI1-Q5_K_M128B82.25 GiB3.09 GiB86.50 GiB2.78 GiB12±22%
Trinity-Large-ThinkingMoEIQ1_M399B84.63 GiB0.75 GiB86.41 GiB2.87 GiB79±37%
Ornith-1.0-397BMoEIQ1_M397B85.09 GiB0.26 GiB86.40 GiB2.88 GiB81±37%
MiniMax-M3MoEIQ1_S427B84.31 GiB1.05 GiB86.38 GiB2.90 GiB66±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XXS236B84.61 GiB0.59 GiB86.24 GiB3.04 GiB65±37%
DeepSeek-V2.5MoEIQ3_XXS236B84.61 GiB0.59 GiB86.24 GiB3.04 GiB65±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XXS236B84.61 GiB0.59 GiB86.24 GiB3.04 GiB65±37%
grok-2MoEUD-IQ1_S270B82.82 GiB2.25 GiB86.22 GiB3.06 GiB22±37%
Codestral-22B-v0.1F3222.2B82.88 GiB1.97 GiB85.91 GiB3.37 GiB12±22%
Mixtral-8x22B-v0.1MoEF32141B82.84 GiB1.97 GiB85.87 GiB3.41 GiB22±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ6_K_L109B83.13 GiB1.69 GiB85.85 GiB3.43 GiB50±37%
DeepSeek-V4-Flash-0731MoEUD-IQ2_M304B84.68 GiB0.02 GiB85.75 GiB3.53 GiB81±37%
DeepSeek-V4-FlashMoEUD-IQ2_M291B84.68 GiB0.02 GiB85.75 GiB3.53 GiB81±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q5_K_S124B83.56 GiB0.77 GiB85.32 GiB3.96 GiB62±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_XXS229B81.91 GiB2.18 GiB85.08 GiB4.20 GiB60±37%
MiniMax-M2.5MoEI1-IQ3_XXS229B81.91 GiB2.18 GiB85.08 GiB4.20 GiB60±37%
Behemoth-X-123B-v2Q5_K_M123B80.55 GiB3.09 GiB84.80 GiB4.48 GiB12±22%
Mistral-Large-Instruct-2411Q5_K_M123B80.55 GiB3.09 GiB84.80 GiB4.48 GiB12±22%
Step-3.5-Flash-REAP-121B-A11BI1-Q5_K_M121B79.79 GiB3.67 GiB84.48 GiB4.80 GiB12±22%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q4_1139B81.21 GiB2.18 GiB84.37 GiB4.91 GiB52±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q4_1139B81.21 GiB2.18 GiB84.37 GiB4.91 GiB52±37%
Mistral-Small-4-119B-2603MoEUD-Q5_K_M119B83.04 GiB0.20 GiB84.27 GiB5.01 GiB74±37%
Qwen3.5-397B-A17BMoEIQ1_S403B82.64 GiB0.26 GiB83.95 GiB5.33 GiB83±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q5_K_M125B82.62 GiB0.21 GiB83.86 GiB5.42 GiB74±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ5_K_M123B82.62 GiB0.21 GiB83.85 GiB5.43 GiB74±37%
Trinity-Large-TrueBaseMoEI1-IQ1_M399B82.07 GiB0.75 GiB83.85 GiB5.43 GiB82±37%
ERNIE-4.5-300B-A47B-PTUD-IQ1_S300B80.54 GiB1.90 GiB83.57 GiB5.71 GiB12±22%
Ace-Step1.5BF16160M82.03 GiB0.43 GiB83.46 GiB5.82 GiB12±22%
Laguna-S-2.1MoEUD-Q5_K_M118B81.83 GiB0.46 GiB83.31 GiB5.97 GiB68±37%
c4ai-command-r-plus-08-2024Q6_K104B79.32 GiB2.25 GiB82.75 GiB6.53 GiB13±22%
Mixtral-8x22B-Instruct-v0.1MoEQ4_K_M141B79.71 GiB1.97 GiB82.74 GiB6.54 GiB23±37%
Mixtral-8x22B-v0.1MoEQ4_K_M141B79.71 GiB1.97 GiB82.74 GiB6.54 GiB23±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ2_K_L236B79.94 GiB1.65 GiB82.63 GiB6.65 GiB51±37%
Qwen3-VL-235B-A22B-InstructMoEQ2_K_L236B79.94 GiB1.65 GiB82.63 GiB6.65 GiB51±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?
2087 of 2118 indexed open-weight models fit a RTX PRO 6000 Blackwell Workstation Edition at 32,768 context with q4_0 KV cache, the largest being xLAM-8x7b-r at BF16. 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.