Model comparison
Qwen2.5-3B-Instruct-abliterated vs Qwen3-8B
At Q4_K_M, Qwen2.5-3B-Instruct-abliterated is the smaller download — 4,034,837,120 bytes against 5,027,783,488. At long context the gap widens: Qwen2.5-3B-Instruct-abliterated's KV cache at 32K is 4.0× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Qwen2.5-3B-Instruct-abliterated | Qwen3-8B | |
|---|---|---|
| Parameters | 3.1B | 8.2B |
| Architecture | qwen2 | qwen3 |
| Layers | 36 | 36 |
| Native context | 32,768 | 40,960 |
| Mixture of experts | no | no |
| Quantizations published | 40 | 51 |
| Smallest quantization | 1.21 GiB | 2.12 GiB |
| Q4_K_M | 3.76 GiB | 4.68 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen2.5-3B-Instruct-abliterated | Qwen3-8B | Ratio |
|---|---|---|---|
| 4,096 | 0.14 GiB | 0.56 GiB | 4.00× |
| 8,192 | 0.28 GiB | 1.13 GiB | 4.00× |
| 16,384 | 0.56 GiB | 2.25 GiB | 4.00× |
| 32,768 | 1.13 GiB | 4.50 GiB | 4.00× |
| 65,536 | 2.25 GiB | 9.00 GiB | 4.00× |
| 131,072 | 4.50 GiB | 18.00 GiB | 4.00× |