How much VRAM does Qwen3.5-4B need?

Q4_K_M is exactly 2,707,513,696 bytes (2.52 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Qwen3.5-4B's KV cache?

1.00 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.

Qwen · vision language

Qwen3.5-4B

Qwen/Qwen3.5-4B

Qwen3.5-4B at Q4_K_M is exactly 2,707,513,696 bytes (2.52 GiB / 2.71 GB) — an effective 4.648 bits per weight, not the nominal 4. Its KV cache at 32K is 1.00 GiB.

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

Parameters

4.7B

Architecture

qwen35

32 layers

Context

262,144

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
UD-IQ2_XXS	1.42 GiB	1,520,217,248	2.610	—	unsloth
UD-IQ2_M	1.64 GiB	1,759,997,088	3.022	—	unsloth
UD-IQ2_M	1.81 GiB	1,942,548,800	3.335	—	unsloth
UD-IQ3_XXS	1.82 GiB	1,949,047,968	3.346	—	unsloth
UD-IQ3_XXS	1.96 GiB	2,103,521,600	3.611	—	unsloth
Q3_K_S	1.96 GiB	2,105,791,648	3.615	—	unsloth
Q3_K_S	2.03 GiB	2,177,505,600	3.738	—	unsloth
Q3_K_M	2.14 GiB	2,293,388,448	3.937	426	unsloth
Q3_K_M	2.21 GiB	2,374,564,160	4.077	441	unsloth
IQ4_XS	2.31 GiB	2,477,053,088	4.253	426	unsloth
IQ4_NL	2.40 GiB	2,579,944,608	4.429	—	unsloth
Q4_0	2.41 GiB	2,583,221,408	4.435	426	unsloth
Q4_K_S	2.41 GiB	2,590,430,368	4.447	—	unsloth
IQ4_XS	2.46 GiB	2,642,012,480	4.536	441	unsloth
Q4_0	2.49 GiB	2,669,209,920	4.582	441	unsloth
Q4_K_S	2.50 GiB	2,683,300,160	4.607	—	unsloth
Q4_K_M	2.52 GiB	2,707,513,696	4.648	426	lmstudio-community
IQ4_NL	2.54 GiB	2,729,175,360	4.685	—	unsloth
Q4_K_M	2.55 GiB	2,740,937,888	4.706	—	unsloth
Q4_1	2.59 GiB	2,784,416,928	4.780	—	unsloth
Q4_K_M	2.64 GiB	2,834,975,040	4.867	441	unsloth
Q4_1	2.68 GiB	2,877,122,880	4.939	—	unsloth
Q5_K_S	2.82 GiB	3,024,934,048	5.193	—	unsloth
Q5_K_S	2.91 GiB	3,124,357,440	5.364	—	unsloth
Q5_K_M	2.93 GiB	3,143,656,608	5.397	—	unsloth
Q5_K_M	2.99 GiB	3,212,790,080	5.516	441	unsloth
Q6_K	3.23 GiB	3,464,055,136	5.947	—	lmstudio-community
Q6_K	3.28 GiB	3,525,956,768	6.053	426	unsloth
Q6_K	3.39 GiB	3,639,654,720	6.248	—	unsloth
Q8_0	4.17 GiB	4,482,402,656	7.695	—	lmstudio-community
Q8_0	4.17 GiB	4,482,403,488	7.695	—	unsloth
Q8_0	4.29 GiB	4,610,580,800	7.915	441	unsloth
BF16	7.85 GiB	8,424,393,632	14.463	—	unsloth
BF16	8.07 GiB	8,665,620,544	14.877	—	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.13 GiB	0.50 GiB	4.00×	8 / 0 / 24
8,192	0.25 GiB	1.00 GiB	4.00×	8 / 0 / 24
16,384	0.50 GiB	2.00 GiB	4.00×	8 / 0 / 24
32,768	1.00 GiB	4.00 GiB	4.00×	8 / 0 / 24
65,536	2.00 GiB	8.00 GiB	4.00×	8 / 0 / 24
131,072	4.00 GiB	16.00 GiB	4.00×	8 / 0 / 24

24 of 32 layers use linear attention, which keeps a fixed-size recurrent state instead of a per-token cache. Those layers do not grow with context at all — treating them as ordinary attention, as a flat formula does, overstates this model's cache by roughly 4.0× at long context.

Compare with

same modality, comparable size

vs Qwen3.5-9B vs Qwen3.5-0.8B vs gemma-4-e4b-it vs Qwen3-VL-4B-Instruct vs Qwen2.5-VL-7B-Instruct vs Qwen3-VL-2B-Instruct 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 2.44 GiB. The real file is 2.52 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

256

Hidden size

2560

Vocab

248,320

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Qwen3.5-4B need?: Q4_K_M is exactly 2,707,513,696 bytes (2.52 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwen3.5-4B's KV cache?: 1.00 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.
Which quantization of Qwen3.5-4B 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.