Is Kimi-VL-A3B-Thinking-2506 or gemma-4-12B-it-qat-q4_0-unquantized smaller?

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

Which uses less memory at long context?

Kimi-VL-A3B-Thinking-2506 — its KV cache at 32K context is 2.6x smaller, because of how its attention layers cache.

Model comparison

Kimi-VL-A3B-Thinking-2506 vs gemma-4-12B-it-qat-q4_0-unquantized

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Kimi-VL-A3B-Thinking-2506's KV cache at 32K is 2.6× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Kimi-VL-A3B-Thinking-2506	gemma-4-12B-it-qat-q4_0-unquantized
Parameters	16.4B	12.0B
Architecture	deepseek2	gemma4
Layers	27	48
Native context	131,072	262,144
Mixture of experts	yes, 64 experts	no
Quantizations published	21	2
Smallest quantization	6.13 GiB	6.50 GiB
Q4_K_M	9.82 GiB	—
Licence	—	apache-2.0

KV cache by context

the term that decides long-context viability

Context	Kimi-VL-A3B-Thinking-2506	gemma-4-12B-it-qat-q4_0-unquantized	Ratio
4,096	0.12 GiB	0.72 GiB	6.06×
8,192	0.24 GiB	0.97 GiB	4.08×
16,384	0.47 GiB	1.47 GiB	3.09×
32,768	0.95 GiB	2.47 GiB	2.60×
65,536	1.90 GiB	4.47 GiB	2.35×
131,072	3.80 GiB	8.47 GiB	2.23×