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. 1952 of 2118 indexed models fit at 8K context with q4_0 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
vision language 173text 1675video 16image 2audio asr 39audio tts 21embedding 26

What fits at 8K context

largest quantization that fits, per model · 1952 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Gemma4-Gutenberg-31BQ4_031.3B16.83 GiB0.68 GiB18.60 GiB0.00 GiB12±22%
Gemma4-Gutenberg-31B-HereticQ4_031.3B16.83 GiB0.68 GiB18.60 GiB0.00 GiB12±22%
Equinox-31BQ4_031.3B16.83 GiB0.68 GiB18.60 GiB0.00 GiB12±22%
gemma-4-31B-it-SDFT-Heretic-RPQ4_030.7B16.83 GiB0.68 GiB18.60 GiB0.00 GiB12±22%
North-Mini-Code-1.0MoEQ4_K_M30.5B17.46 GiB0.15 GiB18.58 GiB0.02 GiB52±37%
GLM-4-32B-0414-Korean-CultureI1-Q4_032.6B17.36 GiB0.13 GiB18.58 GiB0.02 GiB12±22%
GLM-Z1-32B-0414Q4_032.6B17.36 GiB0.13 GiB18.58 GiB0.02 GiB12±22%
GLM-4-32B-0414Q4_032.6B17.36 GiB0.13 GiB18.58 GiB0.02 GiB12±22%
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16UD-IQ3_S33.0B17.53 GiB0.00 GiB18.57 GiB0.03 GiB12±22%
Aurora-Code-1MoEIQ4_XS34.7B17.51 GiB0.04 GiB18.56 GiB0.04 GiB71±37%
grug-35bMoEIQ4_XS35.1B17.51 GiB0.04 GiB18.56 GiB0.04 GiB71±37%
WorldSim-Opus-3.6-35B-A3BMoEIQ4_XS35.1B17.51 GiB0.04 GiB18.56 GiB0.04 GiB71±37%
Qwen3.6-35B-A3B-AnkoMoEIQ4_XS35.1B17.51 GiB0.04 GiB18.56 GiB0.04 GiB71±37%
KAT-Coder-V2.5-DevMoEIQ4_XS34.7B17.51 GiB0.04 GiB18.56 GiB0.04 GiB71±37%
Ornith-1.0-35BMoEIQ4_XS34.7B17.51 GiB0.04 GiB18.56 GiB0.04 GiB71±37%
Nex-N2-miniMoEIQ4_XS35.1B17.51 GiB0.04 GiB18.56 GiB0.04 GiB71±37%
Skywork-R1V3-38BQ4_K_S38.4B17.49 GiB0.00 GiB18.56 GiB0.04 GiB12±22%
Olmo-3.1-32B-InstructQ4_032.2B17.08 GiB0.38 GiB18.56 GiB0.04 GiB12±22%
Olmo-3.1-32B-ThinkQ4_032.2B17.08 GiB0.38 GiB18.56 GiB0.04 GiB12±22%
Olmo-3-32B-ThinkQ4_032.2B17.08 GiB0.38 GiB18.56 GiB0.04 GiB12±22%
Qwen3-Coder-30B-A3B-InstructMoEQ4_K_M30.5B17.35 GiB0.21 GiB18.56 GiB0.04 GiB51±37%
Qwen3-VL-30B-A3B-ThinkingMoEQ4_K_M31.1B17.35 GiB0.21 GiB18.56 GiB0.04 GiB51±37%
MiroThinker-v1.0-30BMoEQ4_K_M30.5B17.35 GiB0.21 GiB18.56 GiB0.04 GiB51±37%
Qwen3-30B-A3BMoEQ4_K_M30.5B17.35 GiB0.21 GiB18.56 GiB0.04 GiB51±37%
Qwen3-30B-A3B-Instruct-2507MoEQ4_K_M30.5B17.35 GiB0.21 GiB18.56 GiB0.04 GiB51±37%
Qwen3-30B-A3B-Thinking-2507MoEQ4_K_M30.5B17.35 GiB0.21 GiB18.56 GiB0.04 GiB51±37%
Pantheon-Proto-RP-1.8-30B-A3BMoEQ4_K_M30.5B17.35 GiB0.21 GiB18.56 GiB0.04 GiB51±37%
Tongyi-DeepResearch-30B-A3BMoEQ4_K_M30.5B17.35 GiB0.21 GiB18.55 GiB0.05 GiB51±37%
spoomplesmaxx-v2.1-30BI1-Q4_128.9B16.88 GiB0.56 GiB18.55 GiB0.05 GiB12±22%
Huihui-granite-4.1-30b-abliteratedI1-Q4_128.9B16.88 GiB0.56 GiB18.55 GiB0.05 GiB12±22%
granite-4.1-30b-hereticI1-Q4_128.9B16.88 GiB0.56 GiB18.55 GiB0.05 GiB12±22%
granite-4.1-30bQ4_128.9B16.88 GiB0.56 GiB18.55 GiB0.05 GiB12±22%
Trinity-2-Codestral-22B-v0.2Q6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
Cydonia-v1.3-Magnum-v4-22BI1-Q6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
Mistral-Small-22B-ArliAI-RPMax-v1.1I1-Q6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
Mistral-Small-Drummer-22BQ6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
magnum-v4-22bI1-Q6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
Mistral-Small-Instruct-2409Q6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
Codestral-22B-v0.1-hfQ6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
Codestral-22B-v0.1Q6_K22.2B17.00 GiB0.49 GiB18.55 GiB0.05 GiB12±22%
gemma-4-31B-itIQ4_NL31.3B16.79 GiB0.68 GiB18.55 GiB0.05 GiB12±22%
dolphin-2.9.1-mixtral-1x22bMoEI1-Q6_K22.2B16.99 GiB0.49 GiB18.55 GiB0.05 GiB7±37%
Mixtral-8x22B-v0.1MoEQ6_K141B16.99 GiB0.49 GiB18.55 GiB0.05 GiB21±37%
reka-flash-3.1I1-Q6_K20.9B17.17 GiB0.29 GiB18.54 GiB0.06 GiB12±22%
reka-flash-3Q6_K20.9B17.17 GiB0.29 GiB18.54 GiB0.06 GiB12±22%
ALIA-40b-fc-2606I1-IQ3_S40.4B17.00 GiB0.42 GiB18.54 GiB0.06 GiB12±22%
ALIA-40b-instruct-2606I1-IQ3_S40.4B17.00 GiB0.42 GiB18.54 GiB0.06 GiB12±22%
granite-4.0-h-smallMoEQ4_032.2B17.51 GiB0.04 GiB18.53 GiB0.07 GiB34±37%
Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q4_K_M30.0B17.11 GiB0.41 GiB18.53 GiB0.07 GiB34±37%
Pantheon-Reasoning-26B-A4B-1.1MoEQ5_K_S26.5B17.36 GiB0.17 GiB18.51 GiB0.09 GiB12±22%
Tess-4-27BQ4_K_M27.8B17.30 GiB0.14 GiB18.51 GiB0.09 GiB12±22%
Yi-34B-200K-DARE-megamerge-v8I1-Q3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
dolphin-2.9.1-yi-1.5-34b-hereticQ3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
dolphin-2.9.1-yi-1.5-34bI1-Q3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
OrionStar-Yi-34B-Chat-LlamaI1-Q3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
Yi-34B-200K-LlamafiedI1-Q3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
Yi-1.5-34BQ3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
Nous-Hermes-2-Yi-34BQ3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
Merged-RP-Stew-V2-34BI1-Q3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 GiB12±22%
Capybara-Tess-Yi-34B-200KQ3_K_L34.4B16.89 GiB0.53 GiB18.51 GiB0.09 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?
1952 of 2118 indexed open-weight models fit a RTX 4000 Ada Generation at 8,192 context with q4_0 KV cache, the largest being Gemma4-Gutenberg-31B at Q4_0. 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.