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. 1816 of 2118 indexed models fit at 32K 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
image 2text 1542vision language 170video 16audio asr 39embedding 26audio tts 21

What fits at 32K context

largest quantization that fits, per model · 1816 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Janus-Pro-7BI1-IQ3_XXS7.4B2.57 GiB15.00 GiB18.60 GiB0.00 GiB12±22%
deepseek-coder-7b-instruct-v1.5I1-IQ3_XXS6.9B2.57 GiB15.00 GiB18.60 GiB0.00 GiB12±22%
Qwen3.6-35B-A3B-uncensored-hereticMoEQ3_K_L35.1B16.97 GiB0.63 GiB18.59 GiB0.01 GiB54±37%
Nex-N2-mini-ultra-uncensored-hereticMoEQ3_K_L35.1B16.97 GiB0.63 GiB18.59 GiB0.01 GiB54±37%
KAT-Coder-V2.5-DevMoEUD-IQ4_XS34.7B16.96 GiB0.63 GiB18.59 GiB0.01 GiB54±37%
Qwen3.6-35B-A3BMoEUD-IQ4_XS36.0B16.96 GiB0.63 GiB18.59 GiB0.01 GiB54±37%
Qwen3-Coder-REAP-25B-A3BMoEQ4_124.9B14.59 GiB3.00 GiB18.58 GiB0.02 GiB24±37%
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16UD-IQ3_S33.0B17.53 GiB0.00 GiB18.57 GiB0.03 GiB12±22%
gemma-4-E2B-it-Uncensored-MAXF325.1B17.33 GiB0.25 GiB18.57 GiB0.03 GiB12±22%
Hy-MT2-30B-A3BMoEQ3_K_L30.1B14.57 GiB3.00 GiB18.57 GiB0.03 GiB25±37%
deepseek-coder-6.7B-kexerI1-IQ1_M6.7B1.54 GiB16.00 GiB18.56 GiB0.04 GiB12±22%
Magicoder-S-DS-6.7BI1-IQ1_M6.7B1.54 GiB16.00 GiB18.56 GiB0.04 GiB12±22%
deepseek-coder-6.7b-baseI1-IQ1_M6.7B1.54 GiB16.00 GiB18.56 GiB0.04 GiB12±22%
Seed-OSS-36B-InstructUD-IQ2_XXS36.2B9.46 GiB8.00 GiB18.56 GiB0.04 GiB12±22%
gemma-4-26B-A4B-itMoEQ4_K_L26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
WizardLM-7B-UncensoredI1-IQ1_M6.7B1.54 GiB16.00 GiB18.56 GiB0.04 GiB12±22%
Llama-2-7B-32K-InstructI1-IQ1_M6.7B1.54 GiB16.00 GiB18.56 GiB0.04 GiB12±22%
Luna-AI-Llama2-UncensoredI1-IQ1_M6.7B1.54 GiB16.00 GiB18.56 GiB0.04 GiB12±22%
Swallow-7b-NVE-instruct-hfI1-IQ1_M6.7B1.54 GiB16.00 GiB18.56 GiB0.04 GiB12±22%
GLM-4.7-Flash-DerestrictedMoEI1-Q4_K_S31.2B15.90 GiB1.65 GiB18.56 GiB0.04 GiB33±37%
Huihui-GLM-4.7-Flash-abliteratedMoEI1-Q4_K_S31.2B15.90 GiB1.65 GiB18.56 GiB0.04 GiB33±37%
Skywork-R1V3-38BQ4_K_S38.4B17.49 GiB0.00 GiB18.56 GiB0.04 GiB12±22%
Goetia-26B-A4B-v1.4MoEI1-Q4_K_M26.0B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
G4-Moonlight-Dusk-26B-A4B-hereticMoEI1-Q4_K_M26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
Pantheon-Reasoning-26B-A4B-1.1-hereticMoEI1-Q4_K_M26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
G4-Moonlight-Dusk-26B-A4BMoEI1-Q4_K_M26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
Chimera-X-26B-A4BMoEI1-Q4_K_M26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
Pantheon-Reasoning-26B-A4B-1.1MoEI1-Q4_K_M26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
Gemma-4-26B-A4B-StyleTune-V2MoEI1-Q4_K_M26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
Gemma-4-26B-A4B-StyleTuneMoEI1-Q4_K_M26.5B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
gemma-4-26b-a4b-heretic-styletune-v2-headMoEI1-Q4_K_M25.8B16.03 GiB1.54 GiB18.56 GiB0.04 GiB12±22%
deepseek-math-7b-instructQ2_K6.9B2.53 GiB15.00 GiB18.56 GiB0.04 GiB12±22%
SambaLingo-Japanese-ChatI1-IQ1_S6.9B1.52 GiB16.00 GiB18.55 GiB0.05 GiB12±22%
NVIDIA-Nemotron-Nano-12B-v2Q6_K_L12.3B9.72 GiB7.75 GiB18.54 GiB0.06 GiB12±22%
dolphin-2.9.3-mistral-7B-32kF167.2B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.3-ParasiteF167.2B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.3-JbliteratedF167.2B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-Instruct-v0.3F167.2B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-v0.3F167.2B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
Mistral-7B-v0.3-Chinese-ChatF167.2B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
mistral-7b-v0.3-bnb-4bitBF167.5B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
Mathstral-7B-v0.1F167.2B13.50 GiB4.00 GiB18.54 GiB0.06 GiB12±22%
Qwen3.6-27B-A3B-CoderMoEI1-Q5_K_S26.7B16.91 GiB0.63 GiB18.54 GiB0.06 GiB46±37%
OpenChat-3.5-7B-Qwen-v2.0KV unresolvedF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
ContextualKunoichi_KTO-7BF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
mistral-7b-uncensoredKV unresolvedF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
xLAM-7b-rBF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
Yarn-Mistral-7b-128kKV unresolvedF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
MegaBeam-Mistral-7B-512kF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
Mistral-7B-Instruct-v0.2BF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
Ninja-v1-RP-WIPKV unresolvedF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
Silicon-Maid-7BKV unresolvedF167.2B13.49 GiB4.00 GiB18.53 GiB0.07 GiB12±22%
c4ai-command-r-08-2024Q2_K_L32.3B12.40 GiB5.00 GiB18.52 GiB0.08 GiB12±22%
phi-4Q6_K14.7B11.20 GiB6.25 GiB18.51 GiB0.09 GiB12±22%
Phi-4-reasoningQ6_K14.7B11.20 GiB6.25 GiB18.51 GiB0.09 GiB12±22%
Phi-4-reasoning-plusQ6_K14.7B11.20 GiB6.25 GiB18.51 GiB0.09 GiB12±22%
TildeOpen-30B-Instruct-LVI1-IQ2_M30.7B9.92 GiB7.50 GiB18.51 GiB0.09 GiB12±22%
Gemma-4-Gembrain-X-Core-31BI1-IQ3_XXS31.3B11.25 GiB6.17 GiB18.50 GiB0.10 GiB12±22%
Gemma-4-Gembrain-X-31BI1-IQ3_XXS31.3B11.25 GiB6.17 GiB18.50 GiB0.10 GiB12±22%
Gemma-4-31B-Isometry-Fabled-PersonaI1-IQ3_XXS31.3B11.25 GiB6.17 GiB18.50 GiB0.10 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?
1816 of 2118 indexed open-weight models fit a RTX 4000 Ada Generation at 32,768 context with f16 KV cache, the largest being Janus-Pro-7B at I1-IQ3_XXS. 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.