NVIDIA · workstation

RTX 2000 Ada Generation

RTX 2000 Ada Generation has 16 GB of VRAM at 224 GB/s — about 14.88 GiB usable after driver and compositor overhead. 1400 of 2118 indexed models fit at 64K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
16 GB
GDDR6
Bandwidth
224 GB/s
128-bit bus
Tensor FP16
48 TF
dense
TDP
70 W
$649 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1163vision language 136embedding 26video 15audio tts 21audio asr 38image 1

What fits at 64K context

largest quantization that fits, per model · 1400 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.7-FlashMoEQ2_K31.2B10.57 GiB3.30 GiB14.88 GiB0.00 GiB16±37%
AMALIA-9B-0626-DPOQ2_K9.2B3.35 GiB10.50 GiB14.88 GiB0.00 GiB9±22%
Olmo-3-7B-InstructQ4_K_M7.3B4.16 GiB9.69 GiB14.88 GiB0.00 GiB9±22%
Olmo-3-7B-ThinkI1-Q4_K_M7.3B4.16 GiB9.69 GiB14.88 GiB0.00 GiB9±22%
Pantheon-Reasoning-26B-A4B-1.1MoEQ2_K_L26.5B11.10 GiB2.79 GiB14.88 GiB0.00 GiB9±22%
Bonsai-8B-unpackedQ4_K_M8.2B4.84 GiB9.00 GiB14.87 GiB0.01 GiB9±22%
Mistral-7B-v0.3Q6_K_L7.2B5.83 GiB8.00 GiB14.87 GiB0.01 GiB9±22%
medgemma-27b-itI1-IQ2_S28.8B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
gemma-3-27b-it-abliterated-refined-visionI1-IQ2_S27.4B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
Nidum-Gemma-3-27B-it-UncensoredI1-IQ2_S27.4B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
gemma-3-27b-it-abliteratedIQ2_S27.4B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
AtomicGPT-gemma3-27bI1-IQ2_S27.4B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
Unbound-v1.12.0-27BI1-IQ2_S27.4B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
Mira-v1.12-Ties-27BI1-IQ2_S27.4B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
gemma-3-27b-itIQ2_S27.4B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
Medgamma27BI1-IQ2_S27.0B8.18 GiB5.61 GiB14.87 GiB0.01 GiB9±22%
granite-20b-code-instruct-8kQ5_K_M20.1B13.79 GiB0.00 GiB14.87 GiB0.01 GiB9±22%
granite-20b-code-base-8kI1-Q5_K_M20.1B13.79 GiB0.00 GiB14.87 GiB0.01 GiB9±22%
granite-34b-code-base-8kI1-IQ3_S33.7B13.79 GiB0.00 GiB14.87 GiB0.01 GiB9±22%
Falcon3-10B-InstructI1-IQ3_XXS10.3B3.80 GiB10.00 GiB14.86 GiB0.02 GiB9±22%
Skywork-R1V3-38BIQ3_M38.4B13.79 GiB0.00 GiB14.86 GiB0.02 GiB9±22%
Huihui-Qwen3.5-35B-A3B-abliteratedMoEI1-IQ3_XXS36.0B12.60 GiB1.25 GiB14.86 GiB0.02 GiB30±37%
Qwen3.5-35B-A3B-BaseMoEI1-IQ3_XXS36.0B12.60 GiB1.25 GiB14.86 GiB0.02 GiB30±37%
Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-DistilledMoEI1-IQ3_XXS36.0B12.60 GiB1.25 GiB14.86 GiB0.02 GiB30±37%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoEIQ4_NL23.6B12.60 GiB1.25 GiB14.86 GiB0.02 GiB28±37%
gemma-4-12BQ6_K_M12.0B9.34 GiB4.47 GiB14.86 GiB0.02 GiB9±22%
Pantheon-Reasoning-27BIQ2_S27.8B9.79 GiB4.00 GiB14.85 GiB0.03 GiB9±22%
Qwen3.5-27BIQ2_S27.8B9.79 GiB4.00 GiB14.85 GiB0.03 GiB9±22%
Salience-1.5-FlashMoEIQ2_XXS31.1B7.85 GiB6.00 GiB14.85 GiB0.03 GiB11±37%
Qwen3-14BUD-IQ1_M14.8B3.79 GiB10.00 GiB14.85 GiB0.03 GiB9±22%
Aurora-Code-1MoEI1-IQ3_M34.7B12.59 GiB1.25 GiB14.84 GiB0.04 GiB30±37%
ZAYA1-8B-CoderMoEQ8_08.8B8.83 GiB5.00 GiB14.84 GiB0.04 GiB9±22%
Nemotron-3-Embed-8B-BF16Q5_K_M8.0B5.30 GiB8.50 GiB14.84 GiB0.04 GiB9±22%
Ministral-3-14B-Instruct-2512UD-IQ2_XXS13.9B3.78 GiB10.00 GiB14.83 GiB0.05 GiB9±22%
Ministral-3-14B-Reasoning-2512UD-IQ2_XXS13.9B3.78 GiB10.00 GiB14.83 GiB0.05 GiB9±22%
Qwen3.5-35B-A3BMoEQ2_K36.0B12.58 GiB1.25 GiB14.83 GiB0.05 GiB30±37%
Qwen3.6-35B-A3BMoEQ2_K36.0B12.58 GiB1.25 GiB14.83 GiB0.05 GiB30±37%
Qwen3.6-27B-Heretic2-Uncensored-Finetune-ThinkingIQ2_M27.4B9.77 GiB4.00 GiB14.83 GiB0.05 GiB9±22%
LFM2-24B-A2BMoEIQ4_NL23.8B12.56 GiB1.25 GiB14.82 GiB0.06 GiB26±37%
granite-4.0-h-smallMoEIQ3_XS32.2B12.83 GiB1.00 GiB14.82 GiB0.06 GiB21±37%
GLM-4.7-Flash-REAP-23B-A3BMoEQ3_K_M23.0B10.50 GiB3.30 GiB14.81 GiB0.07 GiB15±37%
North-Mini-Code-1.0MoEUD-IQ3_S30.5B11.89 GiB1.94 GiB14.81 GiB0.07 GiB22±37%
Qwopus3.6-27B-CoderIQ2_M27.8B9.74 GiB4.00 GiB14.80 GiB0.08 GiB9±22%
Anubis-Mini-8B-v1Q5_K_L8.0B5.76 GiB8.00 GiB14.80 GiB0.08 GiB9±22%
SmolVLM-InstructQ8_02.2B1.80 GiB12.00 GiB14.79 GiB0.09 GiB9±22%
SmolLM2-1.7B-Instruct-UncensoredQ8_01.8B1.80 GiB12.00 GiB14.79 GiB0.09 GiB9±22%
HomunculusIQ2_XS12.5B3.74 GiB10.00 GiB14.79 GiB0.09 GiB9±22%
Apertus-8B-Instruct-2509Q5_K_L8.1B5.72 GiB8.00 GiB14.79 GiB0.09 GiB9±22%
Hy-MT2-7BQ6_K8.0B5.74 GiB8.00 GiB14.78 GiB0.10 GiB9±22%
Hunyuan-7B-InstructQ6_K7.5B5.74 GiB8.00 GiB14.78 GiB0.10 GiB9±22%
Trinity-MiniMoEQ3_K_L26.1B12.66 GiB1.12 GiB14.78 GiB0.10 GiB28±37%
llm-jp-4-8b-instructQ5_K_M8.6B5.73 GiB8.00 GiB14.77 GiB0.11 GiB9±22%
granite-8b-code-base-4kQ4_18.1B4.73 GiB9.00 GiB14.77 GiB0.11 GiB9±22%
Dolphin3.0-Qwen2.5-3bF323.1B11.50 GiB2.25 GiB14.76 GiB0.12 GiB9±22%
VibeThinker-3BF323.1B11.50 GiB2.25 GiB14.76 GiB0.12 GiB9±22%
Qwen2.5-3B-InstructF323.1B11.50 GiB2.25 GiB14.76 GiB0.12 GiB9±22%
Qwen2.5-3BF323.1B11.50 GiB2.25 GiB14.76 GiB0.12 GiB9±22%
Huihui-gemma-3n-E4B-it-abliteratedF167.8B12.80 GiB0.93 GiB14.76 GiB0.12 GiB9±22%
gemma-3n-E4B-itF167.8B12.80 GiB0.93 GiB14.76 GiB0.12 GiB9±22%
Mistral-7B-v0.1KV unresolvedIQ4_NL7.2B5.72 GiB8.00 GiB14.76 GiB0.12 GiB9±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
Prompt processing1954.52 tok/s593.35–2240.4112
Image generation9.32 it/s9.06–9.546
Text generation50.65 tok/s50.60–50.706
Benchmarked· n=12

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 2000 Ada Generation run?
1400 of 2118 indexed open-weight models fit a RTX 2000 Ada Generation at 65,536 context with f16 KV cache, the largest being GLM-4.7-Flash at Q2_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX 2000 Ada Generation 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 2000 Ada Generation fast for local AI?
Its memory bandwidth is 224 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.