How much VRAM does Impish_LLAMA_4B need?

Q4_K_M is exactly 2,778,282,656 bytes (2.59 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Impish_LLAMA_4B's KV cache?

4.00 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.

SicariusSicariiStuff · text

Impish_LLAMA_4B

SicariusSicariiStuff/Impish_LLAMA_4B

Impish_LLAMA_4B at Q4_K_M is exactly 2,778,282,656 bytes (2.59 GiB / 2.78 GB) — an effective 4.925 bits per weight, not the nominal 4. Its KV cache at 32K is 4.00 GiB.

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

Parameters

4.5B

Architecture

llama

32 layers

Context

131,072

native (config.json)

License

llama3.1

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ2_M	1.60 GiB	1,722,633,536	3.054	—	bartowski
Q2_K	1.71 GiB	1,839,738,176	3.261	—	bartowski
IQ3_XXS	1.75 GiB	1,880,116,544	3.333	—	bartowski
IQ3_XS	1.89 GiB	2,027,603,264	3.594	—	bartowski
Q3_K_S	1.96 GiB	2,101,527,200	3.725	—	SicariusSicariiStuff
Q3_K_S	1.96 GiB	2,101,527,872	3.725	—	bartowski
IQ3_M	2.03 GiB	2,183,415,104	3.871	—	bartowski
Q2_K_L	2.07 GiB	2,224,506,176	3.943	—	bartowski
Q3_K_M	2.14 GiB	2,296,562,336	4.071	—	SicariusSicariiStuff
Q3_K_M	2.14 GiB	2,296,563,008	4.071	—	bartowski
Q3_K_L	2.30 GiB	2,464,858,784	4.370	—	SicariusSicariiStuff
Q3_K_L	2.30 GiB	2,464,859,456	4.370	—	bartowski
IQ4_XS	2.36 GiB	2,535,546,176	4.495	—	bartowski
Q4_0	2.47 GiB	2,655,599,936	4.708	—	bartowski
IQ4_NL	2.48 GiB	2,661,104,960	4.718	—	bartowski
Q4_K_S	2.48 GiB	2,664,250,016	4.723	—	SicariusSicariiStuff
Q4_K_S	2.48 GiB	2,664,250,688	4.723	—	bartowski
Q4_K_M	2.59 GiB	2,778,282,656	4.925	—	SicariusSicariiStuff
Q4_K_M	2.59 GiB	2,778,283,328	4.925	—	bartowski
Q4_1	2.71 GiB	2,905,931,072	5.152	—	bartowski
Q4_K_L	2.86 GiB	3,070,707,008	5.444	—	bartowski
Q5_K_S	2.95 GiB	3,163,339,424	5.608	—	SicariusSicariiStuff
Q5_K_S	2.95 GiB	3,163,340,096	5.608	—	bartowski
Q5_K_M	3.01 GiB	3,230,186,144	5.726	—	SicariusSicariiStuff
Q5_K_M	3.01 GiB	3,230,186,816	5.726	—	bartowski
Q5_K_L	3.23 GiB	3,473,360,192	6.157	—	bartowski
Q6_K	3.46 GiB	3,710,333,600	6.577	—	SicariusSicariiStuff
Q6_K	3.46 GiB	3,710,334,272	6.577	—	bartowski
Q6_K_L	3.63 GiB	3,901,179,200	6.916	—	bartowski
Q8_0	4.47 GiB	4,803,216,032	8.515	—	SicariusSicariiStuff
Q8_0	4.47 GiB	4,803,216,704	8.515	—	bartowski
BF16	8.41 GiB	9,033,729,056	16.015	—	bartowski

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.50 GiB	0.50 GiB	—	32 / 0 / 0
8,192	1.00 GiB	1.00 GiB	—	32 / 0 / 0
16,384	2.00 GiB	2.00 GiB	—	32 / 0 / 0
32,768	4.00 GiB	4.00 GiB	—	32 / 0 / 0
65,536	8.00 GiB	8.00 GiB	—	32 / 0 / 0
131,072	16.00 GiB	16.00 GiB	—	32 / 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 2.36 GiB. The real file is 2.59 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 Impish_LLAMA_4B need?: Q4_K_M is exactly 2,778,282,656 bytes (2.59 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Impish_LLAMA_4B's KV cache?: 4.00 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 Impish_LLAMA_4B 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.