NVIDIA · workstation

RTX A6000

RTX A6000 has 48 GB of VRAM at 768 GB/s — about 44.64 GiB usable after driver and compositor overhead. 1978 of 2118 indexed models fit at 64K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
48 GB
GDDR6
Bandwidth
768 GB/s
384-bit bus
Tensor FP16
155 TF
dense
TDP
300 W
$4649 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1693vision language 181image 2audio asr 39audio tts 21video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 1978 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
codegeex4-all-9bIQ2_S9.4B3.51 GiB40.00 GiB44.56 GiB0.08 GiB10±22%
glm-4-9b-chatIQ2_S9.4B3.51 GiB40.00 GiB44.55 GiB0.09 GiB10±22%
Huihui-Qwen3-Coder-Next-abliteratedMoEIQ4_NL79.7B42.04 GiB1.50 GiB44.53 GiB0.11 GiB51±37%
Rocinante-XL-16B-v1BF1616.1B29.93 GiB13.50 GiB44.48 GiB0.16 GiB10±22%
Melody1437-27BQ5_K_M27.8B39.34 GiB4.00 GiB44.40 GiB0.24 GiB10±22%
IQuest-Coder-V1-40B-InstructI1-Q4_139.8B23.24 GiB20.00 GiB44.34 GiB0.30 GiB10±22%
Hypernova-60B-2605MoEI1-Q5_K_M58.7B41.29 GiB2.02 GiB44.30 GiB0.34 GiB38±37%
Apertus-70B-Instruct-2509UD-IQ2_M70.6B23.12 GiB20.00 GiB44.30 GiB0.34 GiB10±22%
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16MoEQ4_K_S75.4B43.15 GiB0.00 GiB44.22 GiB0.42 GiB90±37%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ5_K_M27.8B39.03 GiB4.00 GiB44.10 GiB0.54 GiB10±22%
GLM-4.5VMoEI1-IQ1_S108B31.56 GiB11.50 GiB44.09 GiB0.55 GiB16±37%
llm-surgery-dark-arts-gpt-oss-60b-96a12MoEI1-Q5_K_S60.9B41.56 GiB1.52 GiB44.07 GiB0.57 GiB28±37%
Qwen3.5-122B-A10B-hereticMoEI1-Q2_K123B41.52 GiB1.50 GiB44.04 GiB0.60 GiB47±37%
Gemma-4-31B-Isometry-RPQ8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
Prosopon-31BQ8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
Gemma-4-Novelist-Eclipse-31BQ8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
Giftige-Blume-31B-v1-StyleSwapQ8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
G4-MeroMero-31B-StyleSwapQ8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
Gemma-4-31B-StyleTune-heretic-araQ8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
Pantheon-Reasoning-31B-1.1Q8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
Gemma-4-31B-StyleTuneQ8_032.7B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
Barcenas-StyleTune-31B-FableQ8_032.1B31.79 GiB11.17 GiB44.04 GiB0.60 GiB10±22%
ALIA-40b-fc-2606I1-Q6_K40.4B30.90 GiB12.00 GiB44.01 GiB0.63 GiB10±22%
ALIA-40b-instruct-2606I1-Q6_K40.4B30.90 GiB12.00 GiB44.01 GiB0.63 GiB10±22%
Qwen3.5-122B-A10BMoEIQ2_M125B41.39 GiB1.50 GiB43.92 GiB0.72 GiB47±37%
Maenad-70BI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Rombos-LLM-70b-Llama-3.3I1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
L3.3-Electra-R1-70bI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Llama-3.3_70_b_uncensored_continuedI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Llama-3.3-70B-Instruct-abliteratedI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
grok-oss-Revenant-70BI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
L3.3-70B-Euryale-v2.3I1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Hermes-4-70B-hereticI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Llama-3.1-70BQ2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Golem-70B-v1bI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Legion-V2.1-LLaMa-70BI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Assistant_Pepe_70BI1-Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Athene-70BQ2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Hermes-3-Llama-3.1-70BQ2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
L3.3-70B-Magnum-DiamondQ2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Meta-Llama-3-70B-Instruct-abliterated-v3.5Q2_K_S70.6B22.79 GiB20.00 GiB43.92 GiB0.72 GiB10±22%
Janus-Pro-7BF167.4B12.88 GiB30.00 GiB43.90 GiB0.74 GiB10±22%
c4ai-command-r-plus-08-2024IQ2_XXS104B26.65 GiB16.00 GiB43.83 GiB0.81 GiB10±22%
DeepSeek-R1-Distill-Llama-70BUD-IQ2_M70.6B22.70 GiB20.00 GiB43.82 GiB0.82 GiB10±22%
Mistral-Small-4-119B-2603MoEUD-IQ3_S119B41.36 GiB1.41 GiB43.80 GiB0.84 GiB48±37%
Hermes-4-70BUD-IQ2_M70.6B22.63 GiB20.00 GiB43.75 GiB0.89 GiB10±22%
Llama-3.3-70B-InstructUD-IQ2_M70.6B22.63 GiB20.00 GiB43.75 GiB0.89 GiB10±22%
Hunyuan-A13B-InstructMoEIQ3_M80.4B34.72 GiB8.00 GiB43.71 GiB0.93 GiB10±22%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ2_XS109B30.68 GiB12.00 GiB43.71 GiB0.93 GiB16±37%
deepseek-llm-67b-chatI1-IQ2_XS67.4B18.78 GiB23.75 GiB43.63 GiB1.01 GiB10±22%
deepseek-llm-67b-baseI1-IQ2_XS67.4B18.78 GiB23.75 GiB43.63 GiB1.01 GiB10±22%
openbuddy-deepseek-67b-v15.3-4kI1-IQ2_XS67.4B18.78 GiB23.75 GiB43.63 GiB1.01 GiB10±22%
L3.3-70B-Magnum-v4-SEIQ2_M70.6B22.46 GiB20.00 GiB43.59 GiB1.05 GiB10±22%
Anubis-70B-v1.2IQ2_M70.6B22.46 GiB20.00 GiB43.59 GiB1.05 GiB10±22%
Tess-R1-Limerick-Llama-3.1-70BIQ2_M70.6B22.46 GiB20.00 GiB43.59 GiB1.05 GiB10±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
Image generation14.32 it/s10.40–19.3794
Prompt processing4456.64 tok/s3150.67–5004.8414
Text generation137.32 tok/s131.86–140.2210
Benchmarked· n=94

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 vladmandic-sd-data-benchmark, which publishes no licence — so we display and link rather than redistribute them.

Questions people ask

What AI models can a RTX A6000 run?
1978 of 2118 indexed open-weight models fit a RTX A6000 at 65,536 context with f16 KV cache, the largest being codegeex4-all-9b at IQ2_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX A6000 actually have?
Its nameplate is 48 GB, but about 44.64 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX A6000 fast for local AI?
Its memory bandwidth is 768 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.