Qwen3.6-27B + 3-critic harness matches frontier quality at lower cost
A 27B model with critics beats frontier models? Here's how.
Deep Dive
A developer running Qwen3.6-27B (8-bit) through a 3-critic harness (code review, test review, Playwright e2e) reports that the critics reliably catch the model's extra mistakes, and final output quality after critics is indistinguishable from a frontier run. The plan for this run was written by GLM5.2, and the author guesses the ideal split is a frontier model for planning + Qwen3.6 for execution—using a cheaper model for high-volume implementation where the harness catches errors.
Key Points
- 3-critic harness (code review, test review, Playwright e2e) catches errors from smaller models
- Final quality after critics is indistinguishable from frontier models
- Optimal split: frontier for planning (GLM5.2) + Qwen3.6 for cheap execution
Why It Matters
Enables teams to use small, cheap models for coding at scale without losing quality.