How much VRAM does Qwen3-16B-A3B need?

Q4_K_M is exactly 9,754,940,096 bytes (9.08 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Qwen3-16B-A3B's KV cache?

3.00 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 Qwen3-16B-A3B a mixture-of-experts model?

Yes — 64 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.

kalomaze · text · mixture of experts

Qwen3-16B-A3B

kalomaze/Qwen3-16B-A3B

Qwen3-16B-A3B at Q4_K_M is exactly 9,754,940,096 bytes (9.08 GiB / 9.75 GB) — an effective 4.868 bits per weight, not the nominal 4. Its KV cache at 32K is 3.00 GiB.

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

Parameters

16.0B

total, not active

Architecture

qwen3moe

48 layers

Context

40,960

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ2_XXS	4.12 GiB	4,428,015,648	2.210	—	bartowski
UD-IQ1_S	4.54 GiB	4,873,638,592	2.432	—	unsloth
IQ2_XS	4.59 GiB	4,931,332,128	2.461	—	bartowski
IQ2_S	4.67 GiB	5,013,266,464	2.502	—	bartowski
UD-IQ1_M	4.89 GiB	5,251,044,032	2.621	—	unsloth
UD-IQ2_XXS	5.22 GiB	5,603,300,032	2.796	—	unsloth
IQ2_M	5.24 GiB	5,623,537,696	2.807	—	bartowski
UD-IQ2_M	5.45 GiB	5,854,958,272	2.922	—	unsloth
Q2_K	5.48 GiB	5,881,774,112	2.935	—	bartowski
Q2_K	5.58 GiB	5,986,370,240	2.987	—	unsloth
Q2_K_L	5.64 GiB	6,059,299,520	3.024	—	unsloth
Q2_K_L	5.76 GiB	6,185,646,112	3.087	—	bartowski
IQ3_XXS	6.08 GiB	6,529,507,360	3.259	—	bartowski
IQ3_XS	6.35 GiB	6,822,492,192	3.405	—	bartowski
UD-IQ3_XXS	6.39 GiB	6,861,656,768	3.424	—	unsloth
Q3_K_S	6.56 GiB	7,038,761,664	3.513	—	unsloth
Q3_K_S	6.68 GiB	7,174,420,512	3.580	—	bartowski
IQ3_M	6.99 GiB	7,501,969,440	3.744	—	bartowski
Q3_K_M	6.99 GiB	7,502,231,584	3.744	—	bartowski
Q3_K_L	7.18 GiB	7,706,441,760	3.846	—	bartowski
Q3_K_M	7.24 GiB	7,778,663,104	3.882	—	unsloth
IQ4_XS	8.06 GiB	8,652,165,824	4.318	—	unsloth
IQ4_XS	8.13 GiB	8,732,094,496	4.358	—	bartowski
IQ4_NL	8.50 GiB	9,131,889,344	4.557	—	unsloth
Q4_0	8.53 GiB	9,163,346,624	4.573	—	unsloth
Q4_K_S	8.57 GiB	9,201,619,648	4.592	—	unsloth
IQ4_NL	8.58 GiB	9,207,386,144	4.595	—	bartowski
Q4_0	8.66 GiB	9,301,757,984	4.642	—	bartowski
Q4_K_S	8.85 GiB	9,503,608,864	4.743	—	bartowski
Q4_K_M	9.08 GiB	9,754,940,096	4.868	—	unsloth
Q4_K_M	9.16 GiB	9,830,436,896	4.906	—	bartowski
Q4_K_L	9.37 GiB	10,061,379,616	5.021	—	bartowski
Q4_1	9.41 GiB	10,107,638,464	5.044	—	unsloth
Q4_1	9.43 GiB	10,129,657,888	5.055	—	bartowski
Q5_K_S	10.33 GiB	11,089,679,040	5.534	—	unsloth
Q5_K_S	10.35 GiB	11,108,552,736	5.544	—	bartowski
Q5_K_M	10.63 GiB	11,413,885,632	5.696	—	unsloth
Q5_K_M	10.65 GiB	11,432,759,328	5.706	—	bartowski
Q5_K_L	10.83 GiB	11,624,806,432	5.801	—	bartowski
Q6_K	12.27 GiB	13,176,515,264	6.576	—	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.38 GiB	0.38 GiB	—	48 / 0 / 0
8,192	0.75 GiB	0.75 GiB	—	48 / 0 / 0
16,384	1.50 GiB	1.50 GiB	—	48 / 0 / 0
32,768	3.00 GiB	3.00 GiB	—	48 / 0 / 0
65,536	6.00 GiB	6.00 GiB	—	48 / 0 / 0
131,072	12.00 GiB	12.00 GiB	—	48 / 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 8.40 GiB. The real file is 9.08 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

none

SWA period

—

MLA

Experts

Experts per token

use_sliding_window

false

Questions people ask

How much VRAM does Qwen3-16B-A3B need?: Q4_K_M is exactly 9,754,940,096 bytes (9.08 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwen3-16B-A3B's KV cache?: 3.00 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 Qwen3-16B-A3B a mixture-of-experts model?: Yes — 64 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 Qwen3-16B-A3B 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.