How much VRAM does Llama-2-7B-32K-Instruct need?

Q4_K_M is exactly 4,081,004,288 bytes (3.80 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Llama-2-7B-32K-Instruct's KV cache?

16.00 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.

togethercomputer · text

Llama-2-7B-32K-Instruct

togethercomputer/Llama-2-7B-32K-Instruct

Llama-2-7B-32K-Instruct at Q4_K_M is exactly 4,081,004,288 bytes (3.80 GiB / 4.08 GB) — an effective 4.845 bits per weight, not the nominal 4. Its KV cache at 32K is 16.00 GiB.

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

Parameters

6.7B

Architecture

llama

32 layers

Context

32,768

native (config.json)

License

llama2

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
I1-IQ1_S	1.42 GiB	1,528,583,360	1.815	—	mradermacher
I1-IQ1_M	1.54 GiB	1,650,971,840	1.960	—	mradermacher
I1-IQ2_XXS	1.73 GiB	1,854,952,640	2.202	—	mradermacher
I1-IQ2_XS	1.90 GiB	2,034,914,496	2.416	—	mradermacher
I1-IQ2_S	2.05 GiB	2,196,567,232	2.608	—	mradermacher
I1-Q2_K_S	2.16 GiB	2,319,545,536	2.754	—	mradermacher
I1-IQ2_M	2.20 GiB	2,359,751,872	2.802	—	mradermacher
I1-Q2_K	2.36 GiB	2,532,865,216	3.007	—	mradermacher
I1-IQ3_XXS	2.41 GiB	2,585,392,320	3.069	—	mradermacher
I1-IQ3_XS	2.60 GiB	2,796,524,736	3.320	—	mradermacher
Q2_K	2.63 GiB	2,825,940,736	3.355	—	TheBloke
Q3_K_S	2.75 GiB	2,948,304,640	3.500	—	TheBloke
I1-IQ3_S	2.75 GiB	2,948,306,112	3.500	—	mradermacher
I1-Q3_K_S	2.75 GiB	2,948,306,112	3.500	—	mradermacher
I1-IQ3_M	2.90 GiB	3,114,865,856	3.698	—	mradermacher
Q3_K_M	3.07 GiB	3,298,004,736	3.916	—	TheBloke
I1-Q3_K_M	3.07 GiB	3,298,006,208	3.916	—	mradermacher
Q3_K_L	3.35 GiB	3,597,111,040	4.271	—	TheBloke
I1-Q3_K_L	3.35 GiB	3,597,112,512	4.271	—	mradermacher
I1-IQ4_XS	3.37 GiB	3,619,337,408	4.297	—	mradermacher
Q4_0	3.56 GiB	3,825,807,104	4.542	—	TheBloke
I1-IQ4_NL	3.56 GiB	3,825,808,576	4.542	—	mradermacher
I1-Q4_0	3.57 GiB	3,837,080,768	4.556	—	mradermacher
Q4_K_S	3.59 GiB	3,856,740,096	4.579	—	TheBloke
I1-Q4_K_S	3.59 GiB	3,856,741,568	4.579	—	mradermacher
Q4_K_M	3.80 GiB	4,081,004,288	4.845	—	TheBloke
I1-Q4_K_M	3.80 GiB	4,081,005,760	4.845	—	mradermacher
I1-Q4_1	3.95 GiB	4,238,750,912	5.032	—	mradermacher
Q5_K_S	4.33 GiB	4,651,691,776	5.523	—	TheBloke
Q5_0	4.33 GiB	4,651,691,776	5.523	—	TheBloke
I1-Q5_K_S	4.33 GiB	4,651,693,248	5.523	—	mradermacher
Q5_K_M	4.45 GiB	4,783,156,992	5.679	—	TheBloke
I1-Q5_K_M	4.45 GiB	4,783,158,464	5.679	—	mradermacher
Q6_K	5.15 GiB	5,529,194,240	6.564	—	TheBloke
I1-Q6_K	5.15 GiB	5,529,195,712	6.564	—	mradermacher
Q8_0	6.67 GiB	7,161,089,792	8.502	—	TheBloke

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	2.00 GiB	2.00 GiB	—	32 / 0 / 0
8,192	4.00 GiB	4.00 GiB	—	32 / 0 / 0
16,384	8.00 GiB	8.00 GiB	—	32 / 0 / 0
32,768	16.00 GiB	16.00 GiB	—	32 / 0 / 0
65,536	32.00 GiB	32.00 GiB	—	32 / 0 / 0
131,072	64.00 GiB	64.00 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 3.53 GiB. The real file is 3.80 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

4096

Vocab

32,000

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Llama-2-7B-32K-Instruct need?: Q4_K_M is exactly 4,081,004,288 bytes (3.80 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-2-7B-32K-Instruct's KV cache?: 16.00 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-2-7B-32K-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.