How much VRAM does Atomight-V2.1-0.5B-Inference need?

Q4_K_M is exactly 397,809,504 bytes (0.37 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Atomight-V2.1-0.5B-Inference'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.

NovatasticRoScript · text

Atomight-V2.1-0.5B-Inference

NovatasticRoScript/Atomight-V2.1-0.5B-Inference

Atomight-V2.1-0.5B-Inference at Q4_K_M is exactly 397,809,504 bytes (0.37 GiB / 0.40 GB) — an effective 6.442 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

494M

Architecture

qwen2

24 layers

Context

32,768

native (config.json)

License

mit

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
I1-IQ1_S	0.29 GiB	315,831,168	5.114	—	mradermacher
I1-IQ1_M	0.30 GiB	317,976,192	5.149	—	mradermacher
I1-IQ2_XXS	0.30 GiB	321,551,232	5.207	—	mradermacher
I1-IQ2_XS	0.30 GiB	324,411,264	5.253	—	mradermacher
I1-IQ2_S	0.30 GiB	325,739,136	5.275	—	mradermacher
I1-IQ2_M	0.31 GiB	328,599,168	5.321	—	mradermacher
I1-Q2_K_S	0.31 GiB	331,050,624	5.361	—	mradermacher
I1-IQ3_XXS	0.31 GiB	333,706,368	5.404	—	mradermacher
Q3_K_S	0.32 GiB	338,264,928	5.478	—	mradermacher
I1-Q3_K_S	0.32 GiB	338,265,216	5.478	—	mradermacher
Q2_K	0.32 GiB	338,608,992	5.483	—	mradermacher
I1-IQ3_S	0.32 GiB	338,609,280	5.483	—	mradermacher
I1-Q2_K	0.32 GiB	338,609,280	5.483	—	mradermacher
I1-IQ3_XS	0.32 GiB	338,609,280	5.483	—	mradermacher
I1-IQ3_M	0.32 GiB	342,754,176	5.550	—	mradermacher
I1-IQ4_XS	0.33 GiB	349,404,288	5.658	—	mradermacher
IQ4_XS	0.33 GiB	351,446,880	5.691	—	mradermacher
I1-IQ4_NL	0.33 GiB	352,672,896	5.711	—	mradermacher
I1-Q4_0	0.33 GiB	352,973,952	5.716	—	mradermacher
Q3_K_M	0.33 GiB	355,468,128	5.756	—	mradermacher
I1-Q3_K_M	0.33 GiB	355,468,416	5.756	—	mradermacher
Q3_K_L	0.34 GiB	369,359,712	5.981	—	mradermacher
I1-Q3_K_L	0.34 GiB	369,360,000	5.981	—	mradermacher
I1-Q4_1	0.35 GiB	374,520,960	6.065	—	mradermacher
Q4_K_S	0.36 GiB	385,473,376	6.242	—	mradermacher
I1-Q4_K_S	0.36 GiB	385,473,664	6.242	—	mradermacher
Q4_K_M	0.37 GiB	397,809,504	6.442	—	mradermacher
I1-Q4_K_M	0.37 GiB	397,809,792	6.442	—	mradermacher
Q5_K_S	0.38 GiB	412,711,776	6.683	—	mradermacher
I1-Q5_K_S	0.38 GiB	412,712,064	6.683	—	mradermacher
Q5_K_M	0.39 GiB	420,087,648	6.803	—	mradermacher
I1-Q5_K_M	0.39 GiB	420,087,936	6.803	—	mradermacher
Q6_K	0.47 GiB	505,738,080	8.190	—	mradermacher
I1-Q6_K	0.47 GiB	505,738,368	8.190	—	mradermacher
Q8_0	0.49 GiB	531,069,792	8.600	—	mradermacher
F16	0.93 GiB	994,158,432	16.099	—	mradermacher

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.05 GiB	0.05 GiB	—	24 / 0 / 0
8,192	0.09 GiB	0.09 GiB	—	24 / 0 / 0
16,384	0.19 GiB	0.19 GiB	—	24 / 0 / 0
32,768	0.38 GiB	0.38 GiB	—	24 / 0 / 0
65,536	0.75 GiB	0.75 GiB	—	24 / 0 / 0
131,072	1.50 GiB	1.50 GiB	—	24 / 0 / 0

Compare with

same modality, comparable size

vs Llama-3.2-1B-Instruct vs ced-base vs gemma-3-1b-it vs Qwen3-1.7B vs embeddinggemma-300m vs Qwen3-0.6B 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 0.26 GiB. The real file is 0.37 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

896

Vocab

151,936

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

false

Questions people ask

How much VRAM does Atomight-V2.1-0.5B-Inference need?: Q4_K_M is exactly 397,809,504 bytes (0.37 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Atomight-V2.1-0.5B-Inference'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 Atomight-V2.1-0.5B-Inference 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.