Is GPT-NeoX-20B-Erebus or gemma-4-E4B-it smaller?

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

Which uses less memory at long context?

gemma-4-E4B-it — its KV cache at 32K context is 65.2x smaller, because of how its attention layers cache.

Model comparison

GPT-NeoX-20B-Erebus vs gemma-4-E4B-it

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: gemma-4-E4B-it's KV cache at 32K is 65.2× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	GPT-NeoX-20B-Erebus	gemma-4-E4B-it
Parameters	20.6B	8.0B
Architecture	gptneox	gemma4
Layers	44	42
Native context	2,048	131,072
Mixture of experts	no	no
Quantizations published	23	36
Smallest quantization	4.12 GiB	3.30 GiB
Q4_K_M	—	4.64 GiB
Licence	apache-2.0	apache-2.0

KV cache by context

the term that decides long-context viability

Context	GPT-NeoX-20B-Erebus	gemma-4-E4B-it	Ratio
4,096	4.13 GiB	0.12 GiB	33.52×
8,192	8.25 GiB	0.18 GiB	46.42×
16,384	16.50 GiB	0.29 GiB	57.47×
32,768	33.00 GiB	0.51 GiB	65.24×
65,536	66.00 GiB	0.94 GiB	69.96×
131,072	132.00 GiB	1.82 GiB	72.59×