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. 1942 of 2118 indexed models fit at 64K context with q4_0 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 1665vision language 173audio asr 39image 2video 16audio tts 21embedding 26

What fits at 64K context

largest quantization that fits, per model · 1942 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
dolphin-2.6-mixtral-8x7bMoEI1-IQ3_S46.7B19.03 GiB2.25 GiB22.32 GiB0.00 GiB18±37%
Nous-Hermes-2-Mixtral-8x7B-DPOMoEIQ3_S46.7B19.03 GiB2.25 GiB22.32 GiB0.00 GiB18±37%
Mixtral-8x7B-Instruct-v0.1MoEQ3_K_S46.7B19.03 GiB2.25 GiB22.32 GiB0.00 GiB18±37%
xLAM-8x7b-rMoEQ3_K_S46.7B19.03 GiB2.25 GiB22.32 GiB0.00 GiB18±37%
dolphin-2.5-mixtral-8x7bMoEQ3_K_S46.7B19.03 GiB2.25 GiB22.31 GiB0.01 GiB18±37%
dolphin-2.7-mixtral-8x7bMoEQ3_K_L46.7B19.03 GiB2.25 GiB22.31 GiB0.01 GiB18±37%
Mixtral-8x7B-v0.1MoEQ3_K_S46.7B19.03 GiB2.25 GiB22.31 GiB0.01 GiB18±37%
Qwen3-Coder-30B-A3B-InstructMoEQ5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Salience-1.5-FlashMoEI1-Q5_K_S31.1B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Huihui-Qwen3-VL-30B-A3B-Instruct-abliteratedMoEI1-Q5_K_S31.1B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-VL-30B-A3B-InstructMoEQ5_K_S31.1B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-ABLITERATED-UNCENSOREDMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
MiroThinker-v1.0-30BMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-30B-A3B-YOYO-V5MoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Huihui-Qwen3-30B-A3B-Thinking-2507-abliteratedMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Huihui-Qwen3-30B-A3B-Instruct-2507-abliteratedMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-30B-A3B-abliterated-eroticMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-30B-A3B-abliteratedMoEQ5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-30B-A3BMoEQ5_030.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Huihui-Qwen3-Coder-30B-A3B-Instruct-abliteratedMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Qwen3-Coder-30B-A3B-Instruct-RTPurboMoEI1-Q5_K_S30.5B19.63 GiB1.69 GiB22.31 GiB0.01 GiB35±37%
Nemotron-Cascade-2-30B-A3B-heretic-ara-uncensoredMoEI1-Q4_K_S31.6B20.42 GiB0.91 GiB22.31 GiB0.01 GiB46±37%
Nemotron-Cascade-2-30B-A3BMoEI1-Q4_K_S31.6B20.42 GiB0.91 GiB22.31 GiB0.01 GiB46±37%
Phind-CodeLlama-34B-v2Q4_K_S33.7B17.83 GiB3.38 GiB22.30 GiB0.02 GiB12±22%
CodeLlama-34b-instruct-hfQ4_K_S33.7B17.83 GiB3.38 GiB22.30 GiB0.02 GiB12±22%
WizardLM-1.0-Uncensored-CodeLlama-34bQ4_K_S33.7B17.83 GiB3.38 GiB22.30 GiB0.02 GiB12±22%
Phind-CodeLlama-34B-Python-v1Q4_K_S33.7B17.83 GiB3.38 GiB22.30 GiB0.02 GiB12±22%
Magistry-24B-v1.1Q6_K23.6B18.37 GiB2.81 GiB22.30 GiB0.02 GiB12±22%
WizardCoder-Python-34B-V1.0I1-Q4_033.7B17.81 GiB3.38 GiB22.28 GiB0.04 GiB12±22%
GLM-4.7-Flash-hereticMoEQ5_K_L29.9B20.33 GiB0.93 GiB22.27 GiB0.05 GiB41±37%
Apertus-70B-Instruct-2509UD-IQ1_S70.6B15.46 GiB5.63 GiB22.26 GiB0.06 GiB12±22%
Salience-1.5-ProMoEQ4_136.0B20.91 GiB0.35 GiB22.26 GiB0.06 GiB62±37%
Qwable-v1MoEQ4_136.0B20.91 GiB0.35 GiB22.26 GiB0.06 GiB62±37%
T-SearchMoEQ4_136.0B20.91 GiB0.35 GiB22.26 GiB0.06 GiB62±37%
Open_Gpt4_8x7B_v0.1MoEQ3_K_M46.7B18.97 GiB2.25 GiB22.25 GiB0.07 GiB18±37%
Mixtral-8x7B-MoE-RP-StoryMoEQ3_K_M46.7B18.96 GiB2.25 GiB22.25 GiB0.07 GiB18±37%
Voxtral-Small-24B-2507Q6_K_L24.3B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Devstral-Small-2-24B-Instruct-2512Q6_K_L24.0B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Dolphin3.0-R1-Mistral-24BQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Dolphin3.0-Mistral-24BQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Cydonia_VistralQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Dans-PersonalityEngine-V1.2.0-24bQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Dans-PersonalityEngine-V1.3.0-24bQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Devstral-Small-2505Q6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Mistral-Small-3.2-24B-Instruct-2506Q6_K_L24.0B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
MS3.2-PaintedFantasy-v3-24BQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Precog-24B-v1Q6_K_L—18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Magidonia-24B-v4.3Q6_K_L—18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Magidonia-24B-v4.2.0Q6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
MS-2501-DPE-QwQify-v0.1-24BQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
sarvam-mQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Magistral-Small-2506Q6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Cydonia-24B-v4.1Q6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Cydonia-24B-v4Q6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Mistral-Small-3.1-24B-Instruct-2503Q6_K_L24.0B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Cydonia-24B-v4.3Q6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Cydonia-24B-v4.2.0Q6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Mistral-Small-24B-Instruct-2501-abliteratedQ6_K_L23.6B18.32 GiB2.81 GiB22.25 GiB0.07 GiB12±22%
Dolphin-Mistral-24B-Venice-EditionQ6_K_L24.0B18.32 GiB2.81 GiB22.25 GiB0.07 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 4500 Ada Generation run?
1942 of 2118 indexed open-weight models fit a RTX 4500 Ada Generation at 65,536 context with q4_0 KV cache, the largest being dolphin-2.6-mixtral-8x7b at I1-IQ3_S. 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.