MLX 8-bit (8.5 bpw) quantization of the SFT coding-agent model, for Apple-Silicon serving (e.g. oMLX / mlx-lm). Ready to run as-is (~27 GB).

Full-precision bf16 merged weights: chaddy81/qwen3.6-27b-distilled. LoRA adapter only: chaddy81/qwen3.6-27b-distilled-lora.

bf16 LoRA SFT (r=16, alpha=32), 1 epoch / 710 steps, on 5673 teacher-trajectory records (agentic coding: plan → tool calls → verified patch). Merged into the base, then quantized to MLX 8-bit.

Held-out eval (14 tasks × 3; base = untrained Qwen3.6-27B):

The standout: files-modified 11.47 → 1.05 — surgical, minimal fixes instead of codebase-wide sprawl.

Derivative of Qwen3.6-27B; governed by the base model's license.