NVIDIA · workstation

RTX 4000 Ada Generation

RTX 4000 Ada Generation has 20 GB of VRAM at 360 GB/s — about 18.60 GiB usable after driver and compositor overhead. 1029 of 2118 indexed models fit at 128K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
20 GB
GDDR6
Bandwidth
360 GB/s
160-bit bus
Tensor FP16
107 TF
dense
TDP
130 W
$1250 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 812audio asr 38vision language 121video 16audio tts 20image 1embedding 21

What fits at 128K context

largest quantization that fits, per model · 1029 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Le-Chaton-Slim-23BMoEI1-IQ1_S23.3B4.59 GiB13.00 GiB18.60 GiB0.00 GiB8±37%
Voxtral-Mini-3B-2507Q4_K_L4.7B2.58 GiB15.00 GiB18.59 GiB0.01 GiB12±22%
Llama-3.2-3B-Instruct-abliteratedQ8_03.6B3.58 GiB14.00 GiB18.59 GiB0.01 GiB12±22%
Llama-3.2-3B-Instruct-uncensoredQ8_03.6B3.58 GiB14.00 GiB18.59 GiB0.01 GiB12±22%
dolphincoder-starcoder2-15bKV unresolvedI1-Q3_K_M16.0B7.49 GiB10.00 GiB18.58 GiB0.02 GiB12±22%
gemma-4-12B-coder-fable5-composer2.5-v1-abliteratedQ5_K_M12.0B9.07 GiB8.47 GiB18.58 GiB0.02 GiB12±22%
gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5-abliteratedQ5_K_M12.0B9.07 GiB8.47 GiB18.58 GiB0.02 GiB12±22%
Ling-liteMoEQ4_K_L16.8B10.59 GiB7.00 GiB18.58 GiB0.02 GiB14±37%
Phi-4-mini-instruct-abliteratedQ2_K3.8B1.57 GiB16.00 GiB18.58 GiB0.02 GiB12±22%
Phi-4-mini-reasoningQ2_K3.8B1.57 GiB16.00 GiB18.58 GiB0.02 GiB12±22%
Phi-4-mini-instructQ2_K3.8B1.57 GiB16.00 GiB18.58 GiB0.02 GiB12±22%
Marco-Nano-InstructMoEI1-IQ3_S8.0B3.60 GiB14.00 GiB18.58 GiB0.02 GiB8±37%
gemma-4-26B-A4B-itMoEQ3_K_L26.5B12.29 GiB5.29 GiB18.57 GiB0.03 GiB12±22%
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16UD-IQ3_S33.0B17.53 GiB0.00 GiB18.57 GiB0.03 GiB12±22%
Qwen3-16B-A3BMoEQ2_K16.0B5.58 GiB12.00 GiB18.57 GiB0.03 GiB9±37%
Pantheon-Reasoning-27BIQ2_XS27.8B9.50 GiB8.00 GiB18.56 GiB0.04 GiB12±22%
Qwen3.5-27BIQ2_XS27.8B9.50 GiB8.00 GiB18.56 GiB0.04 GiB12±22%
Skywork-R1V3-38BQ4_K_S38.4B17.49 GiB0.00 GiB18.56 GiB0.04 GiB12±22%
Qwen3.6-28BMoEI1-Q4_028.2B15.05 GiB2.50 GiB18.55 GiB0.05 GiB29±37%
Qwen3.5-28BMoEI1-Q4_028.7B15.05 GiB2.50 GiB18.55 GiB0.05 GiB29±37%
gemma-4-19B-A4B-it-INSTRUCT-Heretic-UncensoredMoEI1-Q5_K_S19.0B12.27 GiB5.29 GiB18.55 GiB0.05 GiB12±22%
gemma-4-19B-A4B-it-The-DECKARD-Heretic-Uncensored-ThinkingMoEI1-Q5_K_S19.0B12.27 GiB5.29 GiB18.55 GiB0.05 GiB12±22%
gemma-4-19b-a4b-it-REAP-hereticMoEI1-Q5_K_S19.0B12.27 GiB5.29 GiB18.55 GiB0.05 GiB12±22%
Gemma-4-19BMoEI1-Q5_K_S19.0B12.27 GiB5.29 GiB18.55 GiB0.05 GiB12±22%
dolphin-2.9.3-mistral-7B-32kI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.3-ParasiteI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.3-JbliteratedI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.3IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Mathstral-7B-v0.1IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-v0.3IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
SciPhi-Self-RAG-Mistral-7B-32kKV unresolvedI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
dolphin-2.2.1-mistral-7bKV unresolvedI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
OpenChat-3.5-7B-Qwen-v2.0KV unresolvedI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
openchat-3.5-0106KV unresolvedI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.1KV unresolvedI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.2I1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
ContextualKunoichi_KTO-7BI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
xLAM-7b-rI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Ninja-v1-RP-WIPKV unresolvedI1-IQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Kunoichi-DPO-v2-7BKV unresolvedIQ1_S7.2B1.50 GiB16.00 GiB18.54 GiB0.06 GiB12±22%
Aura-4BI1-IQ2_S4.5B1.51 GiB16.00 GiB18.53 GiB0.07 GiB12±22%
magnum-v2-4bI1-IQ2_S4.5B1.51 GiB16.00 GiB18.53 GiB0.07 GiB12±22%
gemma-3-12b-it-vl-Gemini-3-Pro-Preview-Heretic-Uncensored-ThinkingI1-Q6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
gemma-3-12b-it-vl-Deepseek-v3.1-Heretic-Uncensored-ThinkingI1-Q6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
gemma-3-12b-it-ultra-uncensored-hereticQ6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
gemma-3-12b-it-vl-GLM-4.7-Flash-Heretic-Uncensored-ThinkingI1-Q6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
Floppa-12B-Gemma3-UncensoredI1-Q6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
gemma-3-12b-it-hereticI1-Q6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
gemma-3-12b-it-abliteratedQ6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
gemma-3-12b-it-abliterated-v2Q6_K11.8B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
gemma-3-12b-itQ6_K12.2B9.00 GiB8.47 GiB18.51 GiB0.09 GiB12±22%
Falcon3-7B-InstructQ3_K_M7.5B3.43 GiB14.00 GiB18.50 GiB0.10 GiB12±22%
OLMoE-1B-7B-0924-InstructMoEI1-IQ1_M6.9B1.53 GiB16.00 GiB18.50 GiB0.10 GiB7±37%
Qwen3.5-27B-Engineer-Deckard-GeminiI1-Q2_K27.7B9.43 GiB8.00 GiB18.49 GiB0.11 GiB12±22%
Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensoredI1-Q2_K27.4B9.43 GiB8.00 GiB18.49 GiB0.11 GiB12±22%
Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-ThinkingI1-Q2_K27.4B9.43 GiB8.00 GiB18.49 GiB0.11 GiB12±22%
Huihui-Qwen3.5-27B-abliteratedI1-Q2_K27.8B9.43 GiB8.00 GiB18.49 GiB0.11 GiB12±22%
Qwen3.5-27B-Unredacted-MAXI1-Q2_K27.4B9.43 GiB8.00 GiB18.49 GiB0.11 GiB12±22%
Qwen3.5-27B-hereticI1-Q2_K27.4B9.43 GiB8.00 GiB18.49 GiB0.11 GiB12±22%
Qwen3.5-27B-DerestrictedI1-Q2_K27.8B9.43 GiB8.00 GiB18.49 GiB0.11 GiB12±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.

Questions people ask

What AI models can a RTX 4000 Ada Generation run?
1029 of 2118 indexed open-weight models fit a RTX 4000 Ada Generation at 131,072 context with f16 KV cache, the largest being Le-Chaton-Slim-23B at I1-IQ1_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX 4000 Ada Generation actually have?
Its nameplate is 20 GB, but about 18.60 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX 4000 Ada Generation fast for local AI?
Its memory bandwidth is 360 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.