How much VRAM does LFM2-1.2B need?

Q4_K_M is exactly 730,893,024 bytes (0.68 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is LFM2-1.2B'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.

LiquidAI · text

LFM2-1.2B

LiquidAI/LFM2-1.2B

LFM2-1.2B at Q4_K_M is exactly 730,893,024 bytes (0.68 GiB / 0.73 GB) — an effective 4.996 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 per layer

Parameters

1.2B

Architecture

lfm2

16 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
Q2_K_L	0.45 GiB	483,396,320	3.304	—	unsloth
Q2_K	0.45 GiB	483,396,320	3.304	—	unsloth
Q3_K_S	0.52 GiB	558,156,512	3.815	—	unsloth
IQ3_M	0.53 GiB	573,575,936	3.921	—	Mungert
IQ3_S	0.53 GiB	573,575,936	3.921	—	Mungert
Q3_K_M	0.56 GiB	600,345,312	4.104	—	unsloth
Q3_K_S	0.58 GiB	624,079,616	4.266	—	Mungert
Q3_K_M	0.60 GiB	641,905,408	4.388	—	Mungert
Q4_0	0.62 GiB	661,148,416	4.519	—	Mungert
IQ4_XS	0.62 GiB	663,376,640	4.535	—	Mungert
Q3_K_L	0.63 GiB	674,411,264	4.610	—	Mungert
Q4_0	0.65 GiB	695,749,344	4.756	—	unsloth
Q4_0	0.65 GiB	695,749,568	4.756	—	LiquidAI
IQ4_NL	0.65 GiB	695,751,424	4.756	—	Mungert
Q4_K_S	0.65 GiB	700,467,936	4.788	—	unsloth
Q4_K_S	0.67 GiB	722,096,896	4.936	—	Mungert
Q4_K_M	0.68 GiB	730,893,024	4.996	—	unsloth
Q4_K_M	0.68 GiB	730,893,248	4.996	—	LiquidAI
Q4_1	0.68 GiB	734,286,592	5.019	—	Mungert
Q4_K_M	0.69 GiB	739,144,448	5.053	—	Mungert
Q4_1	0.71 GiB	760,498,912	5.199	—	unsloth
Q4_K_L	0.72 GiB	771,650,304	5.275	—	Mungert
Q5_0	0.75 GiB	807,424,768	5.519	—	Mungert
Q5_K_S	0.77 GiB	825,248,480	5.641	—	unsloth
Q5_K_M	0.79 GiB	843,352,800	5.765	—	unsloth
Q5_K_M	0.79 GiB	843,353,024	5.765	—	LiquidAI
Q5_K_S	0.79 GiB	852,685,568	5.829	—	Mungert
Q5_K_M	0.80 GiB	861,737,728	5.891	—	Mungert
Q5_1	0.82 GiB	880,562,944	6.019	—	Mungert
Q5_K_L	0.83 GiB	894,243,584	6.113	—	Mungert
Q6_K	0.90 GiB	962,841,312	6.582	—	unsloth
Q6_K	0.90 GiB	962,841,536	6.582	—	LiquidAI
Q6_K_M	0.90 GiB	962,843,392	6.582	—	Mungert
Q6_K_L	0.93 GiB	995,349,248	6.804	—	Mungert
Q8_0	1.16 GiB	1,246,251,744	8.519	—	unsloth
Q8_0	1.16 GiB	1,246,251,968	8.519	—	LiquidAI
Q8_0	1.16 GiB	1,246,253,504	8.519	—	Mungert
F16	2.18 GiB	2,343,324,384	16.018	—	unsloth
F16	2.18 GiB	2,343,324,608	16.018	—	LiquidAI
BF16	2.18 GiB	2,343,326,144	16.018	—	Mungert

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 Qwen3-4B vs HyperCLOVAX-SEED-Text-Instruct-1.5B vs Llama-3.2-1B-Instruct vs ced-base vs gemma-3-1b-it vs Qwen3-1.7B 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 0.61 GiB. The real file is 0.68 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 LFM2-1.2B need?: Q4_K_M is exactly 730,893,024 bytes (0.68 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is LFM2-1.2B'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 LFM2-1.2B 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.