NVIDIA · workstation

RTX PRO 6000 Blackwell Max-Q Workstation Edition

RTX PRO 6000 Blackwell Max-Q Workstation Edition has 96 GB of VRAM at 1792 GB/s — about 89.28 GiB usable after driver and compositor overhead. 2086 of 2118 indexed models fit at 32K context with q8_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
441 TF
dense
TDP
300 W
$8565 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1792vision language 190audio tts 21image 2audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2086 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Devstral-2-123B-Instruct-2512Q5_K_M125B82.25 GiB5.84 GiB89.25 GiB0.03 GiB12±22%
Mistral-Medium-3.5-128BI1-Q5_K_M128B82.25 GiB5.84 GiB89.25 GiB0.03 GiB12±22%
XORTRON-NXTXPRTXXLI1-Q5_K_M128B82.25 GiB5.84 GiB89.25 GiB0.03 GiB12±22%
command-a-plus-05-2026-bf16MoEIQ3_XXS219B87.48 GiB0.76 GiB89.24 GiB0.04 GiB51±37%
MiMo-V2.5MoEKV unresolvedUD-IQ1_M311B86.17 GiB1.99 GiB89.21 GiB0.07 GiB59±37%
gpt-oss-120b-Uncensored-xCloudMoEI1-Q5_K_M117B87.49 GiB0.61 GiB89.09 GiB0.19 GiB67±37%
gpt-oss-120b-abliteratedMoEI1-Q5_K_M117B87.49 GiB0.61 GiB89.09 GiB0.19 GiB67±37%
MiniMax-M2.1MoEIQ3_XXS229B83.91 GiB4.12 GiB89.01 GiB0.27 GiB49±37%
MiniMax-M2MoEIQ3_XXS229B83.91 GiB4.12 GiB89.01 GiB0.27 GiB49±37%
Gemma-4-Dark-Gemistry-31BQ6_K32.7B84.44 GiB3.28 GiB88.80 GiB0.48 GiB12±22%
Qwen3.5-122B-A10BMoEUD-Q5_K_M125B87.21 GiB0.40 GiB88.63 GiB0.65 GiB69±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ2_S310B85.29 GiB1.99 GiB88.33 GiB0.95 GiB59±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ3_XXS235B84.17 GiB3.12 GiB88.32 GiB0.96 GiB44±37%
GLM-4.5-Air-REAP-82B-A12BMoEQ8_081.9B84.17 GiB3.05 GiB88.25 GiB1.03 GiB39±37%
grok-2MoEUD-IQ1_S270B82.82 GiB4.25 GiB88.22 GiB1.06 GiB20±37%
Qwen3.5-REAP-212B-A17BMoEIQ3_M212B86.49 GiB0.50 GiB88.04 GiB1.24 GiB65±37%
step-3.5-flashQ3_K_S199B80.06 GiB6.92 GiB88.01 GiB1.27 GiB12±22%
Step-3.5-Flash-REAP-121B-A11BI1-Q5_K_M121B79.79 GiB6.92 GiB87.74 GiB1.54 GiB12±22%
Codestral-22B-v0.1F3222.2B82.88 GiB3.72 GiB87.66 GiB1.62 GiB12±22%
Mixtral-8x22B-v0.1MoEF32141B82.84 GiB3.72 GiB87.62 GiB1.66 GiB21±37%
Behemoth-X-123B-v2Q5_K_M123B80.55 GiB5.84 GiB87.55 GiB1.73 GiB12±22%
Mistral-Large-Instruct-2411Q5_K_M123B80.55 GiB5.84 GiB87.55 GiB1.73 GiB12±22%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ6_K_L109B83.13 GiB3.19 GiB87.35 GiB1.93 GiB44±37%
MiniMax-M3MoEIQ1_S427B84.31 GiB1.99 GiB87.32 GiB1.96 GiB60±37%
Trinity-Large-ThinkingMoEIQ1_M399B84.63 GiB1.42 GiB87.07 GiB2.21 GiB73±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEI1-IQ3_XXS229B81.91 GiB4.12 GiB87.02 GiB2.26 GiB50±37%
MiniMax-M2.5MoEI1-IQ3_XXS229B81.91 GiB4.12 GiB87.02 GiB2.26 GiB50±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ3_XXS236B84.61 GiB1.12 GiB86.77 GiB2.51 GiB62±37%
DeepSeek-V2.5MoEIQ3_XXS236B84.61 GiB1.12 GiB86.77 GiB2.51 GiB62±37%
DeepSeek-Coder-V2-InstructMoEIQ3_XXS236B84.61 GiB1.12 GiB86.77 GiB2.51 GiB62±37%
Ornith-1.0-397BMoEIQ1_M397B85.09 GiB0.50 GiB86.64 GiB2.64 GiB78±37%
Step-3.7-FlashIQ3_XXS201B78.38 GiB6.92 GiB86.33 GiB2.95 GiB12±22%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q4_1139B81.21 GiB4.12 GiB86.31 GiB2.97 GiB44±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-Q4_1139B81.21 GiB4.12 GiB86.31 GiB2.97 GiB44±37%
dots.llm1.instMoEQ3_K_M143B68.74 GiB16.47 GiB86.23 GiB3.05 GiB25±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q5_K_S124B83.56 GiB1.46 GiB86.01 GiB3.27 GiB58±37%
GLM-4.7MoEUD-TQ1_0358B78.69 GiB6.11 GiB85.84 GiB3.44 GiB40±37%
DeepSeek-V4-Flash-0731MoEUD-IQ2_M304B84.68 GiB0.03 GiB85.76 GiB3.52 GiB80±37%
DeepSeek-V4-FlashMoEUD-IQ2_M291B84.68 GiB0.03 GiB85.76 GiB3.52 GiB80±37%
GLM-4.5MoEUD-TQ1_0358B78.54 GiB6.11 GiB85.69 GiB3.59 GiB40±37%
GLM-4.6MoEUD-TQ1_0357B78.36 GiB6.11 GiB85.51 GiB3.77 GiB40±37%
ERNIE-4.5-300B-A47B-PTUD-IQ1_S300B80.54 GiB3.59 GiB85.25 GiB4.03 GiB12±22%
c4ai-command-r-plus-08-2024Q6_K104B79.32 GiB4.25 GiB84.75 GiB4.53 GiB12±22%
Trinity-Large-TrueBaseMoEI1-IQ1_M399B82.07 GiB1.42 GiB84.52 GiB4.76 GiB75±37%
Mixtral-8x22B-Instruct-v0.1MoEQ4_K_M141B79.71 GiB3.72 GiB84.49 GiB4.79 GiB21±37%
Mixtral-8x22B-v0.1MoEQ4_K_M141B79.71 GiB3.72 GiB84.49 GiB4.79 GiB21±37%
Mistral-Small-4-119B-2603MoEUD-Q5_K_M119B83.04 GiB0.37 GiB84.44 GiB4.84 GiB72±37%
Qwen3.5-397B-A17BMoEIQ1_S403B82.64 GiB0.50 GiB84.19 GiB5.09 GiB80±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ2_K_L236B79.94 GiB3.12 GiB84.10 GiB5.18 GiB45±37%
Qwen3-VL-235B-A22B-InstructMoEQ2_K_L236B79.94 GiB3.12 GiB84.10 GiB5.18 GiB45±37%
Qwen3-235B-A22BMoEQ2_K_L235B79.94 GiB3.12 GiB84.10 GiB5.18 GiB45±37%
Qwen3-235B-A22B-Instruct-2507MoEQ2_K_L235B79.94 GiB3.12 GiB84.10 GiB5.18 GiB45±37%
Qwen3-235B-A22B-Thinking-2507MoEQ2_K_L235B79.94 GiB3.12 GiB84.10 GiB5.18 GiB45±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q5_K_M125B82.62 GiB0.40 GiB84.04 GiB5.24 GiB72±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ5_K_M123B82.62 GiB0.40 GiB84.04 GiB5.24 GiB72±37%
CalmeRys-78B-Orpo-v0.1Q8_078.0B77.16 GiB5.71 GiB84.00 GiB5.28 GiB12±22%
calme-2.3-rys-78bQ8_078.0B77.16 GiB5.71 GiB84.00 GiB5.28 GiB12±22%
GLM-4.6-REAP-268B-A32BMoEUD-IQ1_M269B76.78 GiB6.11 GiB83.93 GiB5.35 GiB37±37%
Ace-Step1.5BF16160M82.03 GiB0.82 GiB83.84 GiB5.44 GiB12±22%
Laguna-S-2.1MoEUD-Q5_K_M118B81.83 GiB0.87 GiB83.72 GiB5.56 GiB65±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 processing7623.24 tok/s5788.82–12034.7618
Text generation269.96 tok/s249.96–271.269
Benchmarked· n=18

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 Max-Q Workstation Edition run?
2086 of 2118 indexed open-weight models fit a RTX PRO 6000 Blackwell Max-Q Workstation Edition at 32,768 context with q8_0 KV cache, the largest being Devstral-2-123B-Instruct-2512 at Q5_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX PRO 6000 Blackwell Max-Q 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 Max-Q 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.