How much VRAM does Qwen3-8B need?

Q4_K_M is exactly 5,027,783,488 bytes (4.68 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Qwen3-8B's KV cache?

4.50 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.

Qwen · text

Qwen3-8B

Qwen/Qwen3-8B

Qwen3-8B at Q4_K_M is exactly 5,027,783,488 bytes (4.68 GiB / 5.03 GB) — an effective 4.911 bits per weight, not the nominal 4. Its KV cache at 32K is 4.50 GiB.

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

Parameters

8.2B

Architecture

qwen3

36 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
UD-IQ1_S	2.12 GiB	2,275,379,008	2.222	—	unsloth
UD-IQ1_M	2.23 GiB	2,396,489,536	2.341	—	unsloth
UD-IQ2_XXS	2.43 GiB	2,605,221,696	2.545	—	unsloth
IQ2_M	2.84 GiB	3,051,914,784	2.981	—	bartowski
UD-IQ2_M	2.90 GiB	3,110,897,472	3.038	—	unsloth
Q2_K	3.06 GiB	3,281,733,152	3.205	—	bartowski
Q2_K	3.06 GiB	3,281,733,440	3.205	—	unsloth
IQ3_XXS	3.14 GiB	3,369,633,312	3.291	—	bartowski
UD-IQ3_XXS	3.18 GiB	3,410,265,920	3.331	—	unsloth
Q2_K_L	3.19 GiB	3,427,592,000	3.348	—	unsloth
IQ3_XS	3.38 GiB	3,626,874,400	3.542	—	bartowski
Q3_K_S	3.51 GiB	3,769,611,808	3.682	—	bartowski
Q3_K_S	3.51 GiB	3,769,612,096	3.682	—	unsloth
Q2_K_L	3.62 GiB	3,889,477,152	3.799	—	bartowski
IQ3_M	3.63 GiB	3,896,620,576	3.806	—	bartowski
Q3_K_M	3.84 GiB	4,124,161,568	4.028	—	bartowski
Q3_K_M	3.84 GiB	4,124,161,856	4.028	399	unsloth
Q3_K_L	4.13 GiB	4,431,394,080	4.328	—	lmstudio-community
Q3_K_L	4.13 GiB	4,431,394,336	4.328	—	bartowski
IQ4_XS	4.25 GiB	4,561,839,648	4.456	—	bartowski
IQ4_XS	4.27 GiB	4,581,287,744	4.475	399	unsloth
Q4_0	4.46 GiB	4,787,332,640	4.676	399	bartowski
IQ4_NL	4.46 GiB	4,793,624,096	4.682	—	bartowski
IQ4_NL	4.46 GiB	4,793,624,384	4.682	—	unsloth
Q4_K_S	4.47 GiB	4,802,012,704	4.690	—	bartowski
Q4_K_S	4.47 GiB	4,802,012,992	4.690	—	unsloth
Q4_K_M	4.68 GiB	5,027,783,488	4.911	399	Qwen
Q4_K_M	4.68 GiB	5,027,783,968	4.911	399	lmstudio-community
Q4_K_M	4.68 GiB	5,027,784,224	4.911	—	bartowski
Q4_K_M	4.68 GiB	5,027,784,512	4.911	—	unsloth
Q4_1	4.89 GiB	5,247,755,808	5.126	—	bartowski
Q4_1	4.89 GiB	5,247,756,096	5.126	—	unsloth
Q4_K_L	5.11 GiB	5,489,669,664	5.362	—	bartowski
Q5_0	5.33 GiB	5,720,761,152	5.588	—	Qwen
Q5_K_S	5.33 GiB	5,720,761,888	5.588	—	bartowski
Q5_K_S	5.33 GiB	5,720,762,176	5.588	—	unsloth
Q5_K_M	5.45 GiB	5,851,112,224	5.715	399	Qwen
Q5_K_M	5.45 GiB	5,851,112,992	5.715	—	bartowski
Q5_K_M	5.45 GiB	5,851,113,280	5.715	—	unsloth
Q5_K_L	5.81 GiB	6,235,207,200	6.090	—	bartowski

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.56 GiB	0.56 GiB	—	36 / 0 / 0
8,192	1.13 GiB	1.13 GiB	—	36 / 0 / 0
16,384	2.25 GiB	2.25 GiB	—	36 / 0 / 0
32,768	4.50 GiB	4.50 GiB	—	36 / 0 / 0
65,536	9.00 GiB	9.00 GiB	—	36 / 0 / 0
131,072	18.00 GiB	18.00 GiB	—	36 / 0 / 0

Compare with

same modality, comparable size

vs gemma-4-E4B-it vs Qwen3-4B vs HyperCLOVAX-SEED-Text-Instruct-1.5B vs Llama-3.2-1B-Instruct vs Qwen3-VL-8B-Instruct-abliterated-v1 vs Llama-3.1-8B-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 4.29 GiB. The real file is 4.68 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

4096

Vocab

151,936

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

false

Questions people ask

How much VRAM does Qwen3-8B need?: Q4_K_M is exactly 5,027,783,488 bytes (4.68 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwen3-8B's KV cache?: 4.50 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.
Which quantization of Qwen3-8B 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.