How much VRAM does Llama-3.2-3B-Instruct-uncensored need?

Q4_K_M is exactly 2,241,003,904 bytes (2.09 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-uncensored'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.

chuanli11 · text

Llama-3.2-3B-Instruct-uncensored

chuanli11/Llama-3.2-3B-Instruct-uncensored

Llama-3.2-3B-Instruct-uncensored at Q4_K_M is exactly 2,241,003,904 bytes (2.09 GiB / 2.24 GB) — an effective 4.971 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

3.6B

Architecture

llama

28 layers

Context

131,072

native (config.json)

License

—

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q2_K	1.39 GiB	1,493,217,664	3.312	—	bartowski
Q2_K	1.39 GiB	1,493,217,856	3.312	—	mradermacher
IQ3_XS	1.53 GiB	1,646,086,528	3.651	—	bartowski
IQ3_XS	1.53 GiB	1,646,086,720	3.651	—	mradermacher
Q3_K_S	1.59 GiB	1,712,146,816	3.798	—	bartowski
Q3_K_S	1.59 GiB	1,712,147,008	3.798	—	mradermacher
IQ3_S	1.59 GiB	1,712,147,008	3.798	—	mradermacher
IQ3_M	1.65 GiB	1,768,966,528	3.924	—	bartowski
IQ3_M	1.65 GiB	1,768,966,720	3.924	—	mradermacher
Q3_K_M	1.73 GiB	1,856,457,088	4.118	—	bartowski
Q3_K_M	1.73 GiB	1,856,457,280	4.118	—	mradermacher
Q2_K_L	1.75 GiB	1,877,985,664	4.165	—	bartowski
Q3_K_L	1.85 GiB	1,984,645,504	4.402	—	bartowski
Q3_K_L	1.85 GiB	1,984,645,696	4.402	—	mradermacher
IQ4_XS	1.90 GiB	2,038,423,936	4.521	—	bartowski
IQ4_XS	1.91 GiB	2,050,220,608	4.548	—	mradermacher
Q4_0	2.00 GiB	2,143,535,488	4.755	—	bartowski
Q4_K_S	2.00 GiB	2,149,826,944	4.768	—	bartowski
Q4_K_S	2.00 GiB	2,149,827,136	4.768	—	mradermacher
Q4_K_M	2.09 GiB	2,241,003,904	4.971	—	bartowski
Q4_K_M	2.09 GiB	2,241,004,096	4.971	—	mradermacher
Q4_K_L	2.36 GiB	2,533,427,584	5.619	—	bartowski
Q5_K_S	2.37 GiB	2,540,388,736	5.635	—	bartowski
Q5_K_S	2.37 GiB	2,540,388,928	5.635	—	mradermacher
Q5_K_M	2.41 GiB	2,593,030,528	5.752	—	bartowski
Q5_K_M	2.41 GiB	2,593,030,720	5.752	—	mradermacher
Q5_K_L	2.64 GiB	2,836,203,904	6.291	—	bartowski
Q6_K	2.76 GiB	2,967,058,816	6.581	—	bartowski
Q6_K	2.76 GiB	2,967,059,008	6.581	—	mradermacher
Q6_K_L	2.94 GiB	3,157,903,744	7.004	—	bartowski
Q8_0	3.58 GiB	3,840,526,720	8.518	—	bartowski
Q8_0	3.58 GiB	3,840,526,912	8.518	—	mradermacher
F16	5.99 GiB	6,433,688,384	14.270	—	hungng
F16	6.73 GiB	7,221,692,512	16.018	—	bartowski
F16	6.73 GiB	7,221,692,992	16.018	—	mradermacher

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 gemma-4-E4B-it 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 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 1.89 GiB. The real file is 2.09 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-uncensored need?: Q4_K_M is exactly 2,241,003,904 bytes (2.09 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-uncensored'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-uncensored 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.