Model comparison

Qwen3.5-2B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING vs Qwen3-4B

At Q4_K_M, Qwen3.5-2B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING is the smaller download — 1,274,397,920 bytes against 2,497,280,256. At long context the gap widens: Qwen3.5-2B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING's KV cache at 32K is 12.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

Qwen3.5-2B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKINGQwen3-4B
Parameters2.2B4.0B
Architectureqwen35qwen3
Layers2436
Native context262,14440,960
Mixture of expertsnono
Quantizations published3629
Smallest quantization0.67 GiB1.01 GiB
Q4_K_M1.19 GiB2.33 GiB
Licenceapache-2.0—

KV cache by context

the term that decides long-context viability
ContextQwen3.5-2B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKINGQwen3-4BRatio
4,0960.05 GiB0.56 GiB11.99×
8,1920.09 GiB1.13 GiB12.00×
16,3840.19 GiB2.25 GiB12.00×
32,7680.38 GiB4.50 GiB12.00×
65,5360.75 GiB9.00 GiB12.00×
131,0721.50 GiB18.00 GiB12.00×