How much VRAM does Llama-3.2-1B-Instruct need?

Q4_K_M is exactly 807,690,688 bytes (0.75 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Llama-3.2-1B-Instruct'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.

meta-llama · text

Llama-3.2-1B-Instruct

meta-llama/Llama-3.2-1B-Instruct

Llama-3.2-1B-Instruct at Q4_K_M is exactly 807,690,688 bytes (0.75 GiB / 0.81 GB) — an effective 5.229 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 from mirror (mirror:unsloth/Llama-3.2-1B-Instruct)

Parameters

1.2B

Architecture

llama

16 layers

Context

131,072

native (config.json)

License

—

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
UD-IQ1_S	0.39 GiB	421,617,696	2.729	—	unsloth
UD-IQ1_M	0.41 GiB	439,181,344	2.843	—	unsloth
UD-IQ2_XXS	0.43 GiB	464,330,784	3.006	—	unsloth
UD-IQ2_M	0.50 GiB	536,322,080	3.472	—	unsloth
UD-IQ3_XXS	0.54 GiB	575,381,536	3.725	—	unsloth
Q2_K_L	0.54 GiB	580,874,272	3.760	—	unsloth
Q2_K	0.54 GiB	580,874,272	3.760	—	unsloth
Q3_K_S	0.60 GiB	641,691,680	4.154	—	unsloth
IQ3_M	0.61 GiB	657,289,344	4.255	—	bartowski
Q3_K_M	0.64 GiB	690,843,680	4.472	147	unsloth
Q3_K_L	0.68 GiB	732,520,896	4.742	—	lmstudio-community
Q3_K_L	0.68 GiB	732,524,672	4.742	—	bartowski
IQ4_XS	0.69 GiB	743,141,408	4.811	147	unsloth
IQ4_XS	0.69 GiB	743,141,504	4.811	147	bartowski
IQ4_NL	0.72 GiB	773,025,824	5.004	—	unsloth
Q4_0	0.72 GiB	773,025,824	5.004	147	unsloth
Q4_0	0.72 GiB	773,025,920	5.004	147	bartowski
Q4_K_S	0.72 GiB	775,647,264	5.021	—	unsloth
Q4_K_S	0.72 GiB	775,647,360	5.021	—	bartowski
Q4_K_M	0.75 GiB	807,690,688	5.229	—	lmstudio-community
Q4_K_M	0.75 GiB	807,694,368	5.229	147	unsloth
Q4_K_M	0.75 GiB	807,694,464	5.229	—	bartowski
Q4_1	0.77 GiB	831,746,080	5.384	—	unsloth
Q4_K_L	0.81 GiB	871,309,440	5.640	—	bartowski
Q5_K_S	0.83 GiB	892,563,488	5.778	—	unsloth
Q5_K_S	0.83 GiB	892,563,584	5.778	—	bartowski
Q5_K_M	0.85 GiB	911,503,392	5.901	147	unsloth
Q5_K_M	0.85 GiB	911,503,488	5.901	147	bartowski
Q5_K_L	0.91 GiB	975,118,464	6.312	—	bartowski
Q6_K	0.95 GiB	1,021,796,800	6.615	147	lmstudio-community
Q6_K	0.95 GiB	1,021,800,480	6.615	147	unsloth
Q6_K	0.95 GiB	1,021,800,576	6.615	147	bartowski
Q6_K_L	1.01 GiB	1,085,415,552	7.026	—	bartowski
Q8_0	1.23 GiB	1,321,079,232	8.552	147	lmstudio-community
Q8_0	1.23 GiB	1,321,082,528	8.552	147	unsloth
Q8_0	1.23 GiB	1,321,083,008	8.552	—	bartowski
F16	2.31 GiB	2,479,595,168	16.052	147	unsloth
BF16	2.31 GiB	2,479,595,264	16.052	—	unsloth
F16	2.31 GiB	2,479,595,360	16.052	—	bartowski

KV cache by context

computed per layer

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

Compare with

same modality, comparable size

vs ced-base vs gemma-3-1b-it vs Qwen3-1.7B vs embeddinggemma-300m vs Llama-3.2-3B-Instruct vs Qwen3-0.6B 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 0.65 GiB. The real file is 0.75 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from mirror:unsloth/Llama-3.2-1B-Instruct

Layers

Attention heads

KV heads

Head dim

Hidden size

2048

Vocab

128,256

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Llama-3.2-1B-Instruct need?: Q4_K_M is exactly 807,690,688 bytes (0.75 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-3.2-1B-Instruct'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 Llama-3.2-1B-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.