How much VRAM does Qwen3-15B-A2B-Base need?

Q4_K_M is exactly 9,496,453,792 bytes (8.84 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Qwen3-15B-A2B-Base's KV cache?

1.50 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.

Is Qwen3-15B-A2B-Base a mixture-of-experts model?

Yes — 128 experts with 8 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

TroyDoesAI · text · mixture of experts

Qwen3-15B-A2B-Base

TroyDoesAI/Qwen3-15B-A2B-Base

Qwen3-15B-A2B-Base at Q4_K_M is exactly 9,496,453,792 bytes (8.84 GiB / 9.50 GB) — an effective 4.877 bits per weight, not the nominal 4. Its KV cache at 32K is 1.50 GiB.

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

Parameters

15.6B

total, not active

Architecture

qwen3moe

24 layers

Context

4,096

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q2_K	5.41 GiB	5,810,950,816	2.984	—	mradermacher
Q3_K_S	6.37 GiB	6,843,681,440	3.515	—	mradermacher
Q3_K_M	7.02 GiB	7,540,918,944	3.873	—	mradermacher
Q3_K_L	7.59 GiB	8,147,782,304	4.184	—	mradermacher
IQ4_XS	7.91 GiB	8,491,795,104	4.361	—	mradermacher
Q4_K_S	8.33 GiB	8,946,377,376	4.595	—	mradermacher
Q4_K_M	8.84 GiB	9,496,453,792	4.877	—	mradermacher
Q5_K_S	10.04 GiB	10,777,813,664	5.535	—	mradermacher
Q5_K_M	10.34 GiB	11,100,349,088	5.701	—	mradermacher
Q6_K	11.93 GiB	12,804,487,840	6.576	—	mradermacher
Q8_0	15.44 GiB	16,575,548,064	8.513	—	mradermacher

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.19 GiB	0.19 GiB	—	24 / 0 / 0
8,192	0.38 GiB	0.38 GiB	—	24 / 0 / 0
16,384	0.75 GiB	0.75 GiB	—	24 / 0 / 0
32,768	1.50 GiB	1.50 GiB	—	24 / 0 / 0
65,536	3.00 GiB	3.00 GiB	—	24 / 0 / 0
131,072	6.00 GiB	6.00 GiB	—	24 / 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 8.16 GiB. The real file is 8.84 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

2048

Vocab

151,936

Sliding window

4096

SWA period

—

MLA

Experts

128

Experts per token

use_sliding_window

false

This model declares a sliding window but sets use_sliding_window: false, so the window is not applied. Honouring the field without the flag understates KV for the whole family.

Questions people ask

How much VRAM does Qwen3-15B-A2B-Base need?: Q4_K_M is exactly 9,496,453,792 bytes (8.84 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwen3-15B-A2B-Base's KV cache?: 1.50 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.
Is Qwen3-15B-A2B-Base a mixture-of-experts model?: Yes — 128 experts, 8 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of Qwen3-15B-A2B-Base 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.