Qwen3.5-4B-Pruned-Coder-UD-Q6_K_XL: A pruned + quantization aware fine tuned coding agent
Trimmed 17% of neurons.
Light bench:
| model | size | prefill t/s | decode t/s | code | instruct | prose | tool |
|---|---|---|---|---|---|---|---|
| 4B base | 3.86 GiB | 11.1 | 4.77 | 3.13 | 2.52 | 1.73 | 8.16 |
| 4B-Pruned | 3.65 GiB | 12.2 | 5.51 | 3.51 | 2.81 | 9.32 | 5.82 |
Calibrated on ~63k code/instruct/tool call tokens for 2 CPU hours.