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. 1358 of 2118 indexed models fit at 128K 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
embedding 26vision language 152text 1104audio asr 38audio tts 21image 1video 16

What fits at 128K context

largest quantization that fits, per model · 1358 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
qwen-indic-v1I1-IQ3_S7.6B3.28 GiB18.00 GiB22.31 GiB0.01 GiB12±22%
gemma-4-26B-A4B-itMoEQ4_K_L26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
OLMoE-1B-7B-0924-InstructMoEQ6_K_L6.9B5.34 GiB16.00 GiB22.31 GiB0.01 GiB8±37%
MiMo-VL-7B-RLI1-Q3_K_S8.3B3.28 GiB18.00 GiB22.31 GiB0.01 GiB12±22%
Kuwutu-7B-CYOA-v2I1-Q3_K_S7.6B3.28 GiB18.00 GiB22.31 GiB0.01 GiB12±22%
Qwen3-Zero-Coder-Reasoning-V2-0.8BI1-IQ3_XXS816M0.34 GiB21.00 GiB22.31 GiB0.01 GiB12±22%
EXAONE-4.0-32BIQ3_XS32.0B12.37 GiB8.84 GiB22.31 GiB0.01 GiB12±22%
Gemma-4-12B-StyleTuneQ8_013.0B12.80 GiB8.47 GiB22.31 GiB0.01 GiB12±22%
gemma-4-12b-heretic-styletune-headQ8_012.0B12.80 GiB8.47 GiB22.31 GiB0.01 GiB12±22%
syrian-gemma-12bQ8_013.0B12.80 GiB8.47 GiB22.31 GiB0.01 GiB12±22%
Goetia-26B-A4B-v1.4MoEI1-Q4_K_M26.0B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
G4-Moonlight-Dusk-26B-A4B-hereticMoEI1-Q4_K_M26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
Pantheon-Reasoning-26B-A4B-1.1-hereticMoEI1-Q4_K_M26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
G4-Moonlight-Dusk-26B-A4BMoEI1-Q4_K_M26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
Chimera-X-26B-A4BMoEI1-Q4_K_M26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
Pantheon-Reasoning-26B-A4B-1.1MoEI1-Q4_K_M26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
Gemma-4-26B-A4B-StyleTune-V2MoEI1-Q4_K_M26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
Gemma-4-26B-A4B-StyleTuneMoEI1-Q4_K_M26.5B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
gemma-4-26b-a4b-heretic-styletune-v2-headMoEI1-Q4_K_M25.8B16.03 GiB5.29 GiB22.31 GiB0.01 GiB12±22%
Jan-code-4bQ5_K_L4.4B3.30 GiB18.00 GiB22.31 GiB0.01 GiB12±22%
Muse-Glimmer-30BQ5_K_L29.8B19.50 GiB1.72 GiB22.30 GiB0.02 GiB12±22%
Qwen3-Coder-REAP-25B-A3BMoEIQ3_XXS24.9B9.31 GiB12.00 GiB22.30 GiB0.02 GiB11±37%
Mistral-7B-v0.1KV unresolvedQ3_K_L7.2B5.26 GiB16.00 GiB22.30 GiB0.02 GiB12±22%
legitus-instruct-v1I1-Q5_K_S8.1B5.23 GiB16.00 GiB22.30 GiB0.02 GiB12±22%
Apertus-8B-Instruct-2509I1-Q5_K_S8.1B5.23 GiB16.00 GiB22.30 GiB0.02 GiB12±22%
Qwen3-VL-Embedding-8BQ3_K_S8.1B3.26 GiB18.00 GiB22.29 GiB0.03 GiB12±22%
Qwen3-Embedding-8BQ3_K_S7.6B3.26 GiB18.00 GiB22.29 GiB0.03 GiB12±22%
Ministral-3-8B-Instruct-2512-BF16-abliteratedI1-Q3_K_L8.9B4.25 GiB17.00 GiB22.29 GiB0.03 GiB12±22%
Amaretto-8BI1-Q3_K_L8.9B4.25 GiB17.00 GiB22.29 GiB0.03 GiB12±22%
Anubis-Mini-8B-v1Q5_K_S8.0B5.25 GiB16.00 GiB22.29 GiB0.03 GiB12±22%
Voxtral-Mini-4B-Realtime-2602KV unresolvedF164.4B8.27 GiB13.00 GiB22.28 GiB0.04 GiB12±22%
GrammarCoder-7B-BaseF167.6B14.23 GiB7.00 GiB22.28 GiB0.04 GiB12±22%
Qwen3.6-28BMoEI1-Q5_K_M28.2B18.77 GiB2.50 GiB22.28 GiB0.04 GiB32±37%
Qwen3.5-28BMoEI1-Q5_K_M28.7B18.77 GiB2.50 GiB22.28 GiB0.04 GiB32±37%
Assistant_Pepe_8BQ5_K_S—5.22 GiB16.00 GiB22.26 GiB0.06 GiB12±22%
Foundation-Sec-8B-InstructI1-Q5_K_S8.0B5.22 GiB16.00 GiB22.26 GiB0.06 GiB12±22%
Foundation-Sec-8B-Instruct-hereticI1-Q5_K_S8.0B5.22 GiB16.00 GiB22.26 GiB0.06 GiB12±22%
dolphin-2.9-llama3-8bQ5_08.0B5.22 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
saiga_llama3_8bQ5_08.0B5.22 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Meta-Llama-3-8B-InstructQ5_K_S8.0B5.22 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
MiniCPM-Llama3-V-2_5Q5_08.5B5.22 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Llama3-ChatQA-1.5-8BQ5_08.0B5.22 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Meta-Llama-3-8BQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Llama-3.1-Tulu-3-8BQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Llama-3-Groq-8B-Tool-UseQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
llama3.1-heretic-unsensoredI1-Q5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Dolphin3.0-Llama3.1-8BQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
dolphin-2.9.4-llama3.1-8bQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Dolphin3.0-Llama3.1-8B-abliteratedQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
LLAMA-3_8B_Unaligned_BETAQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Llama-3.1-8BQ5_08.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Meta-Llama-3-8BQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Deepseek-R1-Distill-NSFW-RPv1Q5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
L3.1-Dark-Reasoning-LewdPlay-evo-Hermes-R1-Uncensored-8B-hereticI1-Q5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
L3.1-Dark-Reasoning-LewdPlay-evo-Hermes-R1-Uncensored-8BI1-Q5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Gluon-8BI1-Q5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Llama3.3-8B-Instruct-Thinking-Heretic-Uncensored-Claude-4.5-Opus-High-ReasoningI1-Q5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-ReasoningI1-Q5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
Hypnos-i1-8BQ5_K_S8.0B5.21 GiB16.00 GiB22.25 GiB0.07 GiB12±22%
grok-oss-Apollyon-8B-hereticI1-Q5_K_S8.0B5.21 GiB16.00 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?
1358 of 2118 indexed open-weight models fit a RTX 4500 Ada Generation at 131,072 context with f16 KV cache, the largest being qwen-indic-v1 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.