Is tinygemma3_cifar or bitnet-b1.58-2B-4T smaller?

Both publish quantizations; compare the table for the exact sizes.

Which uses less memory at long context?

tinygemma3_cifar — its KV cache at 32K context is 18.9x smaller, because of how its attention layers cache.

Model comparison

tinygemma3_cifar vs bitnet-b1.58-2B-4T

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: tinygemma3_cifar's KV cache at 32K is 18.9× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	tinygemma3_cifar	bitnet-b1.58-2B-4T
Parameters	39M	850M
Architecture	gemma3	—
Layers	8	30
Native context	131,072	4,096
Mixture of experts	no	no
Quantizations published	1	2
Smallest quantization	0.04 GiB	0.10 GiB
Q4_K_M	—	—
Licence	wtfpl	mit

KV cache by context

the term that decides long-context viability

Context	tinygemma3_cifar	bitnet-b1.58-2B-4T	Ratio
4,096	0.06 GiB	0.29 GiB	4.69×
8,192	0.08 GiB	0.59 GiB	7.59×
16,384	0.09 GiB	1.17 GiB	12.63×
32,768	0.12 GiB	2.34 GiB	18.90×
65,536	0.19 GiB	4.69 GiB	25.13×
131,072	0.31 GiB	9.38 GiB	30.09×