Model comparison

Goetia-26B-A4B-v1.3-Absolute-Heretic-ARA vs Qwen3-30B-A3B-Thinking-2507

At Q4_K_M, Goetia-26B-A4B-v1.3-Absolute-Heretic-ARA is the smaller download — 16,796,019,808 bytes against 18,556,685,824. At long context the gap widens: Goetia-26B-A4B-v1.3-Absolute-Heretic-ARA's KV cache at 32K is 1.9× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

Goetia-26B-A4B-v1.3-Absolute-Heretic-ARAQwen3-30B-A3B-Thinking-2507
Parameters25.8B30.5B
Architecturegemma4qwen3moe
Layers3048
Native context262,144262,144
Mixture of expertsyes, 128 expertsyes, 128 experts
Quantizations published3551
Smallest quantization7.72 GiB7.05 GiB
Q4_K_M15.64 GiB17.28 GiB
Licenceapache-2.0apache-2.0

KV cache by context

the term that decides long-context viability
ContextGoetia-26B-A4B-v1.3-Absolute-Heretic-ARAQwen3-30B-A3B-Thinking-2507Ratio
4,0960.45 GiB0.38 GiB1.20×
8,1920.61 GiB0.75 GiB1.24×
16,3840.92 GiB1.50 GiB1.63×
32,7681.54 GiB3.00 GiB1.94×
65,5362.79 GiB6.00 GiB2.15×
131,0725.29 GiB12.00 GiB2.27×