ACE-Step · audio tts

Ace-Step1.5

ACE-Step/Ace-Step1.5

Ace-Step1.5 at Q4_K_M is exactly 20,633,952,128 bytes (19.22 GiB / 20.63 GB) Its KV cache at 32K is 1.54 GiB, not the 3.00 GiB a flat formula predicts.

From the file· summed from 10 file(s)From the file· KV per layer
Parameters
160M
Architecture
audiocpp
24 layers
Context
32,768
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
Q8_00.04 GiB40,574,4322.032—audio-cpp
Q8_00.06 GiB61,940,7683.102—audio-cpp
F160.08 GiB84,059,1684.209—audio-cpp
Q8_00.12 GiB127,856,7046.402—audio-cpp
Q8_00.12 GiB127,857,1846.402—audio-cpp
Q8_00.12 GiB127,857,4406.403—audio-cpp
Q8_00.12 GiB127,858,2406.403—audio-cpp
Q8_00.12 GiB127,858,3686.403—audio-cpp
Q8_00.16 GiB167,756,9288.401—audio-cpp
Q8_00.16 GiB172,532,2568.640675audio-cpp
Q8_00.16 GiB175,662,7208.796—audio-cpp
Q8_00.18 GiB193,337,9849.681—audio-cpp
BF160.20 GiB219,096,06410.971—audio-cpp
BF160.20 GiB219,096,54410.971—audio-cpp
BF160.20 GiB219,096,80010.971—audio-cpp
BF160.20 GiB219,097,60010.972—audio-cpp
BF160.20 GiB219,097,72810.972—audio-cpp
F160.23 GiB247,915,23212.415—audio-cpp
Q8_00.23 GiB251,748,92812.607—audio-cpp
Q8_00.28 GiB299,066,46414.976—audio-cpp
F160.29 GiB312,784,19615.663—audio-cpp
BF160.31 GiB332,423,04016.646—audio-cpp
F320.38 GiB411,908,67220.627—audio-cpp
F160.42 GiB452,910,08022.680—audio-cpp
Q8_00.42 GiB454,072,83622.738—audio-cpp
F160.43 GiB456,514,81622.860—audio-cpp
Q8_00.85 GiB915,733,744——audio-cpp
Q8_00.87 GiB930,620,256——audio-cpp
Q8_00.97 GiB1,045,334,432——audio-cpp
Q8_01.02 GiB1,093,739,584——audio-cpp
Q8_01.05 GiB1,129,966,496——audio-cpp
Q8_01.07 GiB1,151,272,416——audio-cpp
F161.17 GiB1,254,813,120——audio-cpp
F161.17 GiB1,255,384,320——audio-cpp
F161.17 GiB1,260,188,416——audio-cpp
Q8_01.18 GiB1,272,140,064——audio-cpp
F161.19 GiB1,277,710,880——audio-cpp
Q8_01.26 GiB1,350,288,416——audio-cpp
Q8_01.27 GiB1,368,991,360——audio-cpp
F161.36 GiB1,463,787,680——audio-cpp

KV cache by context

computed per layer — this model uses sliding-window attention
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.22 GiB0.38 GiB1.68×12 / 12 / 0
8,1920.41 GiB0.75 GiB1.83×12 / 12 / 0
16,3840.79 GiB1.50 GiB1.91×12 / 12 / 0
32,7681.54 GiB3.00 GiB1.95×12 / 12 / 0
65,5363.04 GiB6.00 GiB1.98×12 / 12 / 0
131,0726.04 GiB12.00 GiB1.99×12 / 12 / 0

12 of 24 layers cache only a 128-token window rather than the full context, on a period of . Figures assume the default configuration; --swa-full disables the saving entirely.

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 0.08 GiB. The real file is 19.22 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: a flat formula gives 3.00 GiB at 32K context where the real figure is 1.54 GiB, because most of this model's layers cache a fixed window rather than the whole context.

Architecture

from config.json
Layers
24
Attention heads
16
KV heads
8
Head dim
128
Hidden size
2048
Vocab
64,003
Sliding window
128
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
true

Questions people ask

How much VRAM does Ace-Step1.5 need?
Q4_K_M is exactly 20,633,952,128 bytes (19.22 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Ace-Step1.5's KV cache?
1.54 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of Ace-Step1.5 should I use?
Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.