How much VRAM does LFM2-VL-3B need?

Q4_K_M is exactly 1,563,668,256 bytes (1.46 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is LFM2-VL-3B's KV cache?

0.50 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.

LiquidAI · vision language

LFM2-VL-3B

LiquidAI/LFM2-VL-3B

LFM2-VL-3B at Q4_K_M is exactly 1,563,668,256 bytes (1.46 GiB / 1.56 GB) — an effective 4.171 bits per weight, not the nominal 4. Its KV cache at 32K is 0.50 GiB.

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

Parameters

3.0B

Architecture

lfm2

30 layers

Context

128,000

native (config.json)

License

other

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q4_0	1.38 GiB	1,483,107,968	3.956	—	LiquidAI
Q4_K_M	1.46 GiB	1,563,668,256	4.171	—	LiquidAI
Q5_K_M	1.70 GiB	1,828,957,984	4.879	—	LiquidAI
Q6_K	1.97 GiB	2,110,828,320	5.631	—	LiquidAI
Q8_0	2.55 GiB	2,733,010,560	7.290	—	LiquidAI
F16	4.79 GiB	5,141,458,560	13.715	—	LiquidAI
BF16	4.79 GiB	5,141,458,720	13.715	—	LiquidAI

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.06 GiB	0.23 GiB	3.75×	8 / 0 / 22
8,192	0.13 GiB	0.47 GiB	3.75×	8 / 0 / 22
16,384	0.25 GiB	0.94 GiB	3.75×	8 / 0 / 22
32,768	0.50 GiB	1.88 GiB	3.75×	8 / 0 / 22
65,536	1.00 GiB	3.75 GiB	3.75×	8 / 0 / 22
131,072	2.00 GiB	7.50 GiB	3.75×	8 / 0 / 22

22 of 30 layers use linear attention, which keeps a fixed-size recurrent state instead of a per-token cache. Those layers do not grow with context at all — treating them as ordinary attention, as a flat formula does, overstates this model's cache by roughly 3.8× at long context.

Compare with

same modality, comparable size

vs Qwen3.5-9B vs Qwen3.5-4B vs Qwen3.5-0.8B vs gemma-4-E2B-it vs Qwen3-VL-4B-Instruct vs Qwen2.5-VL-7B-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 1.57 GiB. The real file is 1.46 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

Hidden size

2048

Vocab

65,536

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does LFM2-VL-3B need?: Q4_K_M is exactly 1,563,668,256 bytes (1.46 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is LFM2-VL-3B's KV cache?: 0.50 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 LFM2-VL-3B 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.