Model comparison
Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking 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: Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking's KV cache at 32K is 1.6× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking | gemma-4-12B-it-qat-q4_0-unquantized | |
|---|---|---|
| Parameters | 21.3B | 12.0B |
| Architecture | qwen35 | gemma4 |
| Layers | 48 | 48 |
| Native context | 262,144 | 262,144 |
| Mixture of experts | no | no |
| Quantizations published | 34 | 2 |
| Smallest quantization | 5.30 GiB | 6.50 GiB |
| Q4_K_M | 11.97 GiB | — |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-Thinking | gemma-4-12B-it-qat-q4_0-unquantized | Ratio |
|---|---|---|---|
| 4,096 | 0.19 GiB | 0.72 GiB | 3.83× |
| 8,192 | 0.38 GiB | 0.97 GiB | 2.58× |
| 16,384 | 0.75 GiB | 1.47 GiB | 1.96× |
| 32,768 | 1.50 GiB | 2.47 GiB | 1.65× |
| 65,536 | 3.00 GiB | 4.47 GiB | 1.49× |
| 131,072 | 6.00 GiB | 8.47 GiB | 1.41× |