Is gemma-4-26B-A4B-it or UI-TARS-72B-DPO smaller?

gemma-4-26B-A4B-it is smaller at Q4_K_M: 15.64 GiB against 44.16 GiB.

Which uses less memory at long context?

gemma-4-26B-A4B-it — its KV cache at 32K context is 6.5x smaller, because of how its attention layers cache.

Model comparison

gemma-4-26B-A4B-it vs UI-TARS-72B-DPO

At Q4_K_M, gemma-4-26B-A4B-it is the smaller download — 16,796,017,248 bytes against 47,415,712,640. At long context the gap widens: gemma-4-26B-A4B-it's KV cache at 32K is 6.5× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	gemma-4-26B-A4B-it	UI-TARS-72B-DPO
Parameters	26.5B	73.4B
Architecture	gemma4	qwen2vl
Layers	30	80
Native context	262,144	32,768
Mixture of experts	yes, 128 experts	no
Quantizations published	46	26
Smallest quantization	8.99 GiB	22.11 GiB
Q4_K_M	15.64 GiB	44.16 GiB
Licence	apache-2.0	apache-2.0

KV cache by context

the term that decides long-context viability

Context	gemma-4-26B-A4B-it	UI-TARS-72B-DPO	Ratio
4,096	0.45 GiB	1.25 GiB	2.78×
8,192	0.61 GiB	2.50 GiB	4.13×
16,384	0.92 GiB	5.00 GiB	5.45×
32,768	1.54 GiB	10.00 GiB	6.48×
65,536	2.79 GiB	20.00 GiB	7.16×
131,072	5.29 GiB	40.00 GiB	7.56×