NVIDIA · workstation

RTX 4500 Ada Generation

RTX 4500 Ada Generation has 24 GB of VRAM at 432 GB/s — about 22.32 GiB usable after driver and compositor overhead. 1727 of 2118 indexed models fit at 64K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
24 GB
GDDR6
Bandwidth
432 GB/s
192-bit bus
Tensor FP16
159 TF
dense
TDP
210 W
$2250 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1457vision language 167audio asr 39image 1video 16audio tts 21embedding 26

What fits at 64K context

largest quantization that fits, per model · 1727 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Tongyi-DeepResearch-30B-A3BMoEIQ4_XS30.5B15.33 GiB6.00 GiB22.32 GiB0.00 GiB18±37%
GLM-4-32B-0414-Korean-CultureI1-Q4_K_S32.6B17.41 GiB3.81 GiB22.32 GiB0.00 GiB12±22%
GLM-Z1-32B-0414Q4_K_S32.6B17.41 GiB3.81 GiB22.32 GiB0.00 GiB12±22%
GLM-4-32B-0414Q4_K_S32.6B17.41 GiB3.81 GiB22.32 GiB0.00 GiB12±22%
gemma-4-26B-A4B-it-heretic-ara-v2MoEQ5_K_M25.8B18.52 GiB2.79 GiB22.30 GiB0.02 GiB12±22%
stable-code-3bI1-Q3_K_M2.8B1.30 GiB20.00 GiB22.29 GiB0.03 GiB12±22%
rocket-3BQ3_K_M2.8B1.30 GiB20.00 GiB22.29 GiB0.03 GiB12±22%
Gemma-4-31B-Isometry-RPI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Gemma-4-Dark-Gemistry-31BI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Prosopon-31BI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Gemma-4-Novelist-Eclipse-31BI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Giftige-Blume-31B-v1-StyleSwapI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
G4-MeroMero-31B-StyleSwapI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Gemma-4-31B-StyleTune-heretic-araI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Pantheon-Reasoning-31B-1.1I1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Gemma-4-31B-StyleTuneI1-IQ2_S32.7B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
Barcenas-StyleTune-31B-FableI1-IQ2_S32.1B10.02 GiB11.17 GiB22.28 GiB0.04 GiB12±22%
grug-27bQ4_K_L27.4B17.21 GiB4.00 GiB22.27 GiB0.05 GiB12±22%
Carnice-V2-27bQ4_K_L27.4B17.21 GiB4.00 GiB22.27 GiB0.05 GiB12±22%
Fara1.5-27BQ4_K_L27.4B17.21 GiB4.00 GiB22.27 GiB0.05 GiB12±22%
gemma-4-31B-itUD-IQ2_M31.3B10.01 GiB11.17 GiB22.27 GiB0.05 GiB12±22%
Qwen3.5-35B-A3BMoEQ4_K_S36.0B20.01 GiB1.25 GiB22.27 GiB0.05 GiB46±37%
Qwen3.6-35B-A3BMoEQ4_K_S36.0B20.01 GiB1.25 GiB22.27 GiB0.05 GiB46±37%
Qwen3-Coder-30B-A3B-InstructMoEIQ4_XS30.5B15.25 GiB6.00 GiB22.25 GiB0.07 GiB18±37%
Qwen3-VL-30B-A3B-InstructMoEIQ4_XS31.1B15.25 GiB6.00 GiB22.25 GiB0.07 GiB18±37%
Pantheon-Reasoning-26B-A4B-1.1MoEQ5_K_M26.5B18.47 GiB2.79 GiB22.25 GiB0.07 GiB12±22%
Salience-1.5-FlashMoEI1-IQ4_XS31.1B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Huihui-Qwen3-VL-30B-A3B-Instruct-abliteratedMoEI1-IQ4_XS31.1B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-ABLITERATED-UNCENSOREDMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
MiroThinker-v1.0-30BMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Qwen3-30B-A3B-YOYO-V5MoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Huihui-Qwen3-30B-A3B-Thinking-2507-abliteratedMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Huihui-Qwen3-30B-A3B-Instruct-2507-abliteratedMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Qwen3-30B-A3B-abliterated-eroticMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Huihui-Qwen3-Coder-30B-A3B-Instruct-abliteratedMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Qwen3-Coder-30B-A3B-Instruct-RTPurboMoEI1-IQ4_XS30.5B15.24 GiB6.00 GiB22.24 GiB0.08 GiB18±37%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEIQ2_XXS—19.74 GiB1.50 GiB22.23 GiB0.09 GiB45±37%
umt5-xxlF325.7B21.17 GiB0.00 GiB22.22 GiB0.10 GiB12±22%
gemma-2-27b-itIQ2_M27.2B8.75 GiB12.31 GiB22.20 GiB0.12 GiB12±22%
magnum-v4-27bIQ2_M27.2B8.75 GiB12.31 GiB22.20 GiB0.12 GiB12±22%
dolphin-2.6-mixtral-8x7bMoEI1-IQ2_S46.7B13.16 GiB8.00 GiB22.19 GiB0.13 GiB12±37%
xLAM-8x7b-rMoEIQ2_S46.7B13.16 GiB8.00 GiB22.19 GiB0.13 GiB12±37%
Phi-4-reasoningQ4_114.7B8.63 GiB12.50 GiB22.19 GiB0.13 GiB12±22%
Phi-4-reasoning-plusQ4_114.7B8.63 GiB12.50 GiB22.19 GiB0.13 GiB12±22%
phi-4Q4_114.7B8.63 GiB12.50 GiB22.19 GiB0.13 GiB12±22%
Codestral-22B-v0.1Q2_K_S22.2B7.12 GiB14.00 GiB22.18 GiB0.14 GiB12±22%
dolphin-2.9.1-mixtral-1x22bMoEQ2_K_S22.2B7.12 GiB14.00 GiB22.18 GiB0.14 GiB7±37%
NVIDIA-Nemotron-Nano-12B-v2Q3_K_M12.3B5.61 GiB15.50 GiB22.18 GiB0.14 GiB12±22%
GRM-2.6-Plus-0628Q4_K_M27.8B17.12 GiB4.00 GiB22.18 GiB0.14 GiB12±22%
GLM-4.7-Flash-hereticMoEQ4_129.9B17.86 GiB3.30 GiB22.18 GiB0.14 GiB26±37%
Darwin-35B-A3B-OpusMoEQ4_K_M36.0B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
Aurora-Code-1MoEQ4_K_M34.7B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
grug-35b-v2MoEQ4_K_M35.1B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
grug-35bMoEQ4_K_M35.1B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
WorldSim-Opus-3.6-35B-A3BMoEQ4_K_M35.1B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
Qwen3.6-35B-A3B-AnkoMoEQ4_K_M35.1B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
KAT-Coder-V2.5-DevMoEQ4_K_M34.7B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
Ornith-1.0-35BMoEQ4_K_M34.7B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
Nex-N2-miniMoEQ4_K_M35.1B19.92 GiB1.25 GiB22.18 GiB0.14 GiB46±37%
●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 4500 Ada Generation run?
1727 of 2118 indexed open-weight models fit a RTX 4500 Ada Generation at 65,536 context with f16 KV cache, the largest being Tongyi-DeepResearch-30B-A3B at IQ4_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX 4500 Ada Generation actually have?
Its nameplate is 24 GB, but about 22.32 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX 4500 Ada Generation fast for local AI?
Its memory bandwidth is 432 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.