How much VRAM does LFM2.5-230M need?

Q4_K_M is exactly 153,406,048 bytes (0.14 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is LFM2.5-230M's KV cache?

0.38 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 · text

LFM2.5-230M

LiquidAI/LFM2.5-230M

LFM2.5-230M at Q4_K_M is exactly 153,406,048 bytes (0.14 GiB / 0.15 GB) — an effective 5.343 bits per weight, not the nominal 4. Its KV cache at 32K is 0.38 GiB.

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

Parameters

230M

Architecture

lfm2

14 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
UD-IQ2_M	0.10 GiB	105,458,880	3.673	—	unsloth
UD-IQ3_XXS	0.11 GiB	112,884,928	3.932	—	unsloth
Q3_K_S	0.12 GiB	127,495,360	4.441	—	unsloth
Q3_K_M	0.12 GiB	133,754,048	4.659	132	unsloth
IQ4_XS	0.13 GiB	144,002,240	5.016	132	unsloth
Q4_0	0.14 GiB	149,080,928	5.192	—	LiquidAI
IQ4_NL	0.14 GiB	149,081,280	5.192	—	unsloth
Q4_0	0.14 GiB	149,245,120	5.198	132	unsloth
Q4_K_S	0.14 GiB	149,671,104	5.213	—	unsloth
Q4_K_M	0.14 GiB	153,406,048	5.343	—	lazos
Q4_K_M	0.14 GiB	153,406,304	5.343	—	LiquidAI
Q4_K_M	0.14 GiB	153,406,656	5.343	132	unsloth
Q4_1	0.15 GiB	159,239,360	5.546	—	unsloth
Q5_K_S	0.16 GiB	169,397,440	5.900	—	unsloth
Q5_K_M	0.16 GiB	171,625,312	5.978	—	LiquidAI
Q5_K_M	0.16 GiB	171,625,664	5.978	132	unsloth
Q6_K	0.18 GiB	190,982,752	6.652	—	lazos
Q6_K	0.18 GiB	190,983,008	6.652	—	LiquidAI
Q6_K	0.18 GiB	190,983,360	6.652	132	unsloth
Q8_0	0.23 GiB	246,598,240	8.589	—	lazos
Q8_0	0.23 GiB	246,598,496	8.589	—	LiquidAI
Q8_0	0.23 GiB	246,598,848	8.589	132	unsloth
BF16	0.43 GiB	461,884,256	16.087	—	LiquidAI
F16	0.43 GiB	461,884,256	16.087	—	LiquidAI
BF16	0.43 GiB	461,884,384	16.087	—	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.05 GiB	0.11 GiB	2.33×	6 / 0 / 8
8,192	0.09 GiB	0.22 GiB	2.33×	6 / 0 / 8
16,384	0.19 GiB	0.44 GiB	2.33×	6 / 0 / 8
32,768	0.38 GiB	0.88 GiB	2.33×	6 / 0 / 8
65,536	0.75 GiB	1.75 GiB	2.33×	6 / 0 / 8
131,072	1.50 GiB	3.50 GiB	2.33×	6 / 0 / 8

8 of 14 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 2.3× at long context.

Compare with

same modality, comparable size

vs ced-base vs Qwen3-0.6B vs SmolLM2-135M-Instruct vs gemma-3-270m-it vs tinygemma3_cifar vs LFM2.5-350M 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 0.12 GiB. The real file is 0.14 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

1024

Vocab

65,536

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does LFM2.5-230M need?: Q4_K_M is exactly 153,406,048 bytes (0.14 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is LFM2.5-230M's KV cache?: 0.38 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 LFM2.5-230M 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.