How much VRAM does ERNIE-4.5-21B-A3B-PT need?

Q4_K_M is exactly 13,331,015,456 bytes (12.42 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is ERNIE-4.5-21B-A3B-PT's KV cache?

1.75 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

baidu · text

ERNIE-4.5-21B-A3B-PT

baidu/ERNIE-4.5-21B-A3B-PT

ERNIE-4.5-21B-A3B-PT at Q4_K_M is exactly 13,331,015,456 bytes (12.42 GiB / 13.33 GB) — an effective 4.859 bits per weight, not the nominal 4. Its KV cache at 32K is 1.75 GiB.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

21.9B

Architecture

ernie4_5-moe

28 layers

Context

131,072

native (config.json)

License

—

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ2_XS	5.91 GiB	6,346,976,352	2.313	—	bartowski
IQ2_S	5.93 GiB	6,369,913,952	2.322	—	bartowski
UD-TQ1_0	6.06 GiB	6,504,202,016	2.371	—	unsloth
UD-IQ1_S	6.58 GiB	7,066,500,896	2.576	—	unsloth
IQ2_M	6.67 GiB	7,164,537,952	2.611	—	bartowski
UD-IQ1_M	6.74 GiB	7,235,276,576	2.637	—	unsloth
UD-IQ2_XXS	7.22 GiB	7,757,506,336	2.828	—	unsloth
UD-IQ2_M	7.47 GiB	8,025,599,776	2.925	—	unsloth
Q2_K	7.54 GiB	8,091,872,352	2.949	—	bartowski
Q2_K_L	7.59 GiB	8,152,596,256	2.971	—	unsloth
Q2_K	7.59 GiB	8,152,596,256	2.971	—	unsloth
Q2_K_L	7.60 GiB	8,155,995,232	2.973	—	bartowski
IQ3_XXS	8.38 GiB	8,993,094,752	3.278	—	bartowski
IQ3_XS	8.71 GiB	9,350,859,872	3.408	—	bartowski
UD-IQ3_XXS	8.89 GiB	9,540,402,976	3.477	—	unsloth
Q3_K_S	8.98 GiB	9,639,608,096	3.514	—	unsloth
Q3_K_S	9.17 GiB	9,850,039,392	3.590	—	bartowski
IQ3_M	9.59 GiB	10,293,595,232	3.752	—	bartowski
Q3_K_M	9.59 GiB	10,297,855,072	3.753	—	bartowski
Q3_K_M	9.80 GiB	10,524,815,136	3.836	—	unsloth
Q3_K_L	9.93 GiB	10,663,750,464	3.887	—	lmstudio-community
Q3_K_L	9.93 GiB	10,663,750,752	3.887	—	bartowski
IQ4_XS	11.01 GiB	11,823,021,856	4.309	—	unsloth
IQ4_XS	11.14 GiB	11,958,004,832	4.359	—	bartowski
IQ4_NL	11.62 GiB	12,475,596,576	4.547	—	unsloth
Q4_0	11.64 GiB	12,501,360,416	4.557	—	unsloth
Q4_K_S	11.68 GiB	12,536,094,496	4.569	—	unsloth
IQ4_NL	11.74 GiB	12,603,698,272	4.594	—	bartowski
Q4_0	11.90 GiB	12,782,611,552	4.659	—	bartowski
Q4_K_S	12.12 GiB	13,012,642,912	4.743	—	bartowski
Q4_K_M	12.42 GiB	13,331,015,456	4.859	—	unsloth
Q4_K_M	12.57 GiB	13,499,267,904	4.920	—	lmstudio-community
Q4_K_M	12.57 GiB	13,499,268,192	4.920	—	bartowski
Q4_K_L	12.63 GiB	13,563,391,072	4.944	—	bartowski
Q4_1	12.83 GiB	13,778,452,256	5.022	—	unsloth
Q4_1	12.93 GiB	13,881,322,592	5.060	—	bartowski
Q5_K_S	14.10 GiB	15,142,297,376	5.519	—	unsloth
Q5_K_S	14.18 GiB	15,230,340,192	5.551	—	bartowski
Q5_K_M	14.51 GiB	15,585,330,976	5.681	—	unsloth
Q5_K_M	14.67 GiB	15,751,453,792	5.741	—	bartowski

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.22 GiB	0.22 GiB	—	28 / 0 / 0
8,192	0.44 GiB	0.44 GiB	—	28 / 0 / 0
16,384	0.88 GiB	0.88 GiB	—	28 / 0 / 0
32,768	1.75 GiB	1.75 GiB	—	28 / 0 / 0
65,536	3.50 GiB	3.50 GiB	—	28 / 0 / 0
131,072	7.00 GiB	7.00 GiB	—	28 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs gemma-4-12B-it-qat-q4_0-unquantized vs gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 11.50 GiB. The real file is 12.42 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

2560

Vocab

103,424

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does ERNIE-4.5-21B-A3B-PT need?: Q4_K_M is exactly 13,331,015,456 bytes (12.42 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is ERNIE-4.5-21B-A3B-PT's KV cache?: 1.75 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 ERNIE-4.5-21B-A3B-PT 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.