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

Q4_K_M is exactly 2,241,004,000 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-abliterated'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.

huihui-ai · text

Llama-3.2-3B-Instruct-abliterated

huihui-ai/Llama-3.2-3B-Instruct-abliterated

Llama-3.2-3B-Instruct-abliterated at Q4_K_M is exactly 2,241,004,000 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
IQ1_S	0.93 GiB	997,440,256	2.212	—	MaziyarPanahi
I1-IQ1_S	0.93 GiB	997,440,704	2.212	—	mradermacher
IQ1_M	0.98 GiB	1,053,473,536	2.337	—	MaziyarPanahi
I1-IQ1_M	0.98 GiB	1,053,473,984	2.337	—	mradermacher
I1-IQ2_XXS	1.07 GiB	1,146,862,784	2.544	—	mradermacher
IQ2_XS	1.15 GiB	1,229,830,912	2.728	—	MaziyarPanahi
I1-IQ2_XS	1.15 GiB	1,229,831,360	2.728	—	mradermacher
I1-IQ2_S	1.23 GiB	1,323,619,520	2.936	—	mradermacher
I1-IQ2_M	1.30 GiB	1,398,330,560	3.102	—	mradermacher
Q2_K	1.39 GiB	1,493,218,048	3.312	—	MaziyarPanahi
I1-Q2_K	1.39 GiB	1,493,218,496	3.312	—	mradermacher
I1-IQ3_XXS	1.41 GiB	1,518,064,832	3.367	—	mradermacher
IQ3_XS	1.53 GiB	1,646,086,912	3.651	—	MaziyarPanahi
I1-IQ3_XS	1.53 GiB	1,646,087,360	3.651	—	mradermacher
Q3_K_S	1.59 GiB	1,712,147,200	3.798	—	MaziyarPanahi
I1-IQ3_S	1.59 GiB	1,712,147,648	3.798	—	mradermacher
I1-Q3_K_S	1.59 GiB	1,712,147,648	3.798	—	mradermacher
I1-IQ3_M	1.65 GiB	1,768,967,360	3.924	—	mradermacher
Q3_K_M	1.73 GiB	1,856,457,472	4.118	—	MaziyarPanahi
I1-Q3_K_M	1.73 GiB	1,856,457,920	4.118	—	mradermacher
Q3_K_L	1.85 GiB	1,984,645,888	4.402	—	MaziyarPanahi
I1-Q3_K_L	1.85 GiB	1,984,646,336	4.402	—	mradermacher
IQ4_XS	1.90 GiB	2,038,424,320	4.521	—	MaziyarPanahi
I1-IQ4_XS	1.90 GiB	2,038,424,768	4.521	—	mradermacher
I1-Q4_0	2.00 GiB	2,143,536,320	4.755	—	mradermacher
Q4_K_S	2.00 GiB	2,149,827,328	4.769	—	MaziyarPanahi
I1-Q4_K_S	2.00 GiB	2,149,827,776	4.769	—	mradermacher
Q4_K_M	2.09 GiB	2,241,004,000	4.971	—	Hasaranga85
Q4_K_M	2.09 GiB	2,241,004,288	4.971	—	MaziyarPanahi
I1-Q4_K_M	2.09 GiB	2,241,004,736	4.971	—	mradermacher
Q5_K_S	2.37 GiB	2,540,389,120	5.635	—	MaziyarPanahi
I1-Q5_K_S	2.37 GiB	2,540,389,568	5.635	—	mradermacher
Q5_K_M	2.41 GiB	2,593,030,912	5.752	—	MaziyarPanahi
I1-Q5_K_M	2.41 GiB	2,593,031,360	5.752	—	mradermacher
Q6_K	2.76 GiB	2,967,059,200	6.581	—	MaziyarPanahi
I1-Q6_K	2.76 GiB	2,967,059,648	6.581	—	mradermacher
Q8_0	3.58 GiB	3,840,527,104	8.518	—	MaziyarPanahi

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 1.89 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-abliterated need?: Q4_K_M is exactly 2,241,004,000 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-abliterated'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-abliterated 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.