How much VRAM does llama-3.2-3b-instruct need?

Q4_K_M is exactly 2,019,377,248 bytes (1.88 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is llama-3.2-3b-instruct's KV cache?

3.50 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.

unsloth · text

llama-3.2-3b-instruct

unsloth/llama-3.2-3b-instruct

llama-3.2-3b-instruct at Q4_K_M is exactly 2,019,377,248 bytes (1.88 GiB / 2.02 GB) — an effective 5.028 bits per weight, not the nominal 4. Its KV cache at 32K is 3.50 GiB.

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

Parameters

3.2B

Architecture

—

28 layers

Context

131,072

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ2_XS	1.02 GiB	1,100,548,416	2.740	—	juiceb0xc0de
IQ3_M	1.49 GiB	1,599,668,544	3.983	—	juiceb0xc0de
IQ3_M	1.49 GiB	1,599,668,576	3.983	—	Zynerji
Q3_K_M	1.57 GiB	1,687,158,880	4.201	—	juiceb0xc0de
IQ4_XS	1.70 GiB	1,829,110,080	4.555	—	juiceb0xc0de
IQ4_XS	1.70 GiB	1,829,110,112	4.555	—	Zynerji
Q4_0	1.79 GiB	1,917,191,072	4.774	—	mfielding92
Q4_K_M	1.88 GiB	2,019,377,248	5.028	—	juiceb0xc0de
Q4_K_M	1.88 GiB	2,019,377,504	5.028	—	Zynerji
Q5_K_M	2.16 GiB	2,322,153,568	5.782	—	juiceb0xc0de
Q5_K_M	2.16 GiB	2,322,153,824	5.782	—	Zynerji
Q6_K	2.46 GiB	2,643,853,408	6.583	—	juiceb0xc0de
Q6_K	2.46 GiB	2,643,853,664	6.583	—	Zynerji
Q8_0	3.19 GiB	3,421,898,848	8.521	—	juiceb0xc0de
Q8_0	3.19 GiB	3,421,899,104	8.521	—	Zynerji
Q8_0	3.19 GiB	3,421,899,680	8.521	—	mfielding92
Q2_K2 shards	4.08 GiB	4,379,869,504	10.906	—	Sweaterdog
F16	5.99 GiB	6,433,687,648	16.020	—	juiceb0xc0de
Q4_K_M2 shards	6.24 GiB	6,702,444,864	16.690	—	Sweaterdog
Q5_K_M2 shards	7.23 GiB	7,766,978,880	19.340	—	Sweaterdog
Q8_02 shards	10.73 GiB	11,520,418,112	28.687	—	Sweaterdog
F163 shards	34.37 GiB	36,909,386,080	—	—	Sweaterdog

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.44 GiB	0.44 GiB	—	28 / 0 / 0
8,192	0.88 GiB	0.88 GiB	—	28 / 0 / 0
16,384	1.75 GiB	1.75 GiB	—	28 / 0 / 0
32,768	3.50 GiB	3.50 GiB	—	28 / 0 / 0
65,536	7.00 GiB	7.00 GiB	—	28 / 0 / 0
131,072	14.00 GiB	14.00 GiB	—	28 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-4B vs Qwen3-8B 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 1.68 GiB. The real file is 1.88 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

3072

Vocab

128,256

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does llama-3.2-3b-instruct need?: Q4_K_M is exactly 2,019,377,248 bytes (1.88 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is llama-3.2-3b-instruct's KV cache?: 3.50 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 llama-3.2-3b-instruct 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.