Model comparison
Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking
At Q4_K_M, Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking is the smaller download — 12,852,817,792 bytes against 18,556,689,568. At long context the gap widens: Qwen3.6-21B-IQ-Ultra-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-Coder-30B-A3B-Instruct | Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking | |
|---|---|---|
| Parameters | 30.5B | 21.3B |
| Architecture | qwen3moe | qwen35 |
| Layers | 48 | 48 |
| Native context | 262,144 | 262,144 |
| Mixture of experts | yes, 128 experts | no |
| Quantizations published | 46 | 34 |
| Smallest quantization | 7.46 GiB | 5.30 GiB |
| Q4_K_M | 17.28 GiB | 11.97 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3-Coder-30B-A3B-Instruct | Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking | Ratio |
|---|---|---|---|
| 4,096 | 0.38 GiB | 0.19 GiB | 2.00× |
| 8,192 | 0.75 GiB | 0.38 GiB | 2.00× |
| 16,384 | 1.50 GiB | 0.75 GiB | 2.00× |
| 32,768 | 3.00 GiB | 1.50 GiB | 2.00× |
| 65,536 | 6.00 GiB | 3.00 GiB | 2.00× |
| 131,072 | 12.00 GiB | 6.00 GiB | 2.00× |