Model comparison

Qwen3.5-21B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking vs Qwen3-Coder-30B-A3B-Instruct

At Q4_K_M, Qwen3.5-21B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking is the smaller download — 12,852,818,432 bytes against 18,556,689,568. At long context the gap widens: Qwen3.5-21B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking's KV cache at 32K is 2.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

Qwen3.5-21B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingQwen3-Coder-30B-A3B-Instruct
Parameters21.3B30.5B
Architectureqwen35qwen3moe
Layers4848
Native context262,144262,144
Mixture of expertsnoyes, 128 experts
Quantizations published3446
Smallest quantization5.30 GiB7.46 GiB
Q4_K_M11.97 GiB17.28 GiB
Licenceapache-2.0apache-2.0

KV cache by context

the term that decides long-context viability
ContextQwen3.5-21B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingQwen3-Coder-30B-A3B-InstructRatio
4,0960.19 GiB0.38 GiB2.00×
8,1920.38 GiB0.75 GiB2.00×
16,3840.75 GiB1.50 GiB2.00×
32,7681.50 GiB3.00 GiB2.00×
65,5363.00 GiB6.00 GiB2.00×
131,0726.00 GiB12.00 GiB2.00×