How much VRAM does Supra-50M-Base need?

Q4_K_M is exactly 37,445,408 bytes (0.03 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Supra-50M-Base'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.

SupraLabs · text

Supra-50M-Base

SupraLabs/Supra-50M-Base

Supra-50M-Base at Q4_K_M is exactly 37,445,408 bytes (0.03 GiB / 0.04 GB) — an effective 5.785 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

52M

Architecture

llama

12 layers

Context

1,024

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ1_S	0.02 GiB	22,951,072	3.546	—	SupraLabs
IQ1_M	0.02 GiB	23,467,168	3.625	—	SupraLabs
IQ2_XXS	0.02 GiB	24,327,328	3.758	—	SupraLabs
IQ2_XS	0.02 GiB	25,113,760	3.880	—	SupraLabs
TQ1_0	0.02 GiB	25,138,976	3.884	—	SupraLabs
IQ2_S	0.02 GiB	25,715,872	3.973	—	SupraLabs
TQ2_0	0.02 GiB	26,392,352	4.077	—	SupraLabs
IQ2_M	0.02 GiB	26,404,000	4.079	—	SupraLabs
IQ3_XXS	0.03 GiB	27,411,616	4.235	—	SupraLabs
Q2_K_S	0.03 GiB	28,362,912	4.381	—	SupraLabs
Q2_K	0.03 GiB	28,751,648	4.442	—	SupraLabs
IQ3_XS	0.03 GiB	30,159,008	4.659	—	SupraLabs
Q3_K_S	0.03 GiB	30,988,064	4.787	—	SupraLabs
IQ3_S	0.03 GiB	30,988,064	4.787	—	SupraLabs
IQ3_M	0.03 GiB	31,704,864	4.898	—	SupraLabs
Q3_K_M	0.03 GiB	32,728,864	5.056	—	SupraLabs
Q3_K_L	0.03 GiB	33,826,592	5.226	—	SupraLabs
IQ4_XS	0.03 GiB	33,838,880	5.228	—	SupraLabs
Q4_0	0.03 GiB	34,539,296	5.336	—	SupraLabs
IQ4_NL	0.03 GiB	34,674,464	5.357	—	SupraLabs
Q4_K_S	0.03 GiB	35,731,232	5.520	—	SupraLabs
Q4_1	0.03 GiB	36,751,136	5.677	—	SupraLabs
Q4_K_M	0.03 GiB	37,445,408	5.785	—	SupraLabs
Q5_0	0.04 GiB	38,962,976	6.019	—	SupraLabs
Q5_K_S	0.04 GiB	39,503,648	6.103	—	SupraLabs
Q5_K_M	0.04 GiB	40,959,776	6.327	—	SupraLabs
Q5_1	0.04 GiB	41,174,816	6.361	—	SupraLabs
Q6_K	0.04 GiB	45,758,240	7.069	—	SupraLabs
Q8_0	0.05 GiB	56,202,016	8.682	—	SupraLabs
F16	0.10 GiB	104,739,616	16.180	—	SupraLabs
F16	0.10 GiB	104,739,616	16.180	—	SupraLabs
BF16	0.10 GiB	104,739,616	16.180	—	SupraLabs
F32	0.19 GiB	208,286,496	32.176	—	SupraLabs

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	—	12 / 0 / 0
8,192	0.09 GiB	0.09 GiB	—	12 / 0 / 0
16,384	0.19 GiB	0.19 GiB	—	12 / 0 / 0
32,768	0.38 GiB	0.38 GiB	—	12 / 0 / 0
65,536	0.75 GiB	0.75 GiB	—	12 / 0 / 0
131,072	1.50 GiB	1.50 GiB	—	12 / 0 / 0

Compare with

same modality, comparable size

vs tinygemma3_cifar vs rfdetr vs Supra-Router-51M vs FireRedPunc vs Tinystories-gpt-0.1-3m 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 0.03 GiB. The real file is 0.03 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

512

Vocab

32,000

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Supra-50M-Base need?: Q4_K_M is exactly 37,445,408 bytes (0.03 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Supra-50M-Base'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 Supra-50M-Base 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.