Model comparison
Qwen3.6-27B vs Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B
At Q4_K_M, Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B is the smaller download — 11,312,942,112 bytes against 16,547,397,888. At long context the gap widens: Qwen3.6-27B's KV cache at 32K is 1.7× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Qwen3.6-27B | Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B | |
|---|---|---|
| Parameters | 27.8B | 18.4B |
| Architecture | qwen35 | llama |
| Layers | 64 | 28 |
| Native context | 262,144 | 131,072 |
| Mixture of experts | no | yes, 8 experts |
| Quantizations published | 40 | 31 |
| Smallest quantization | 8.74 GiB | 6.56 GiB |
| Q4_K_M | 15.41 GiB | 10.54 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3.6-27B | Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B | Ratio |
|---|---|---|---|
| 4,096 | 0.25 GiB | 0.44 GiB | 1.75× |
| 8,192 | 0.50 GiB | 0.88 GiB | 1.75× |
| 16,384 | 1.00 GiB | 1.75 GiB | 1.75× |
| 32,768 | 2.00 GiB | 3.50 GiB | 1.75× |
| 65,536 | 4.00 GiB | 7.00 GiB | 1.75× |
| 131,072 | 8.00 GiB | 14.00 GiB | 1.75× |