Pi-adapted Qwen 3.8 27B goes 3/3 building an FPS agentically; base model 0/2 with 10x more calls
ben_burtenshaw · x · 2026-11-21
A rigorous controlled test on a MacBook (M5 Pro, 64GB) shows the pi-adapted Qwen 3.8 27B fine-tune actually works: with identical briefs (Tenaz + Pi, Splash 1.7.0) for building a raycaster FPS with no human in the loop, the fine-tune succeeded 3/3 with only 390–520 model calls per run and at most 1 session running out of turns. The base 27B went 0/2 — one run BLOCKED after 9.5h, one COMPLETEWITHGAPS after 6.3h — burning 3,000–4,900 calls with 26–39 sessions exhausting turns as it dumped its output budget into reasoning. Its last game: one pixel of mouse input turned the view 126°. The base model decodes faster (65–69 vs 47–55 tok/s, since the DFlash2 drafter was trained on it), but task completion was night and day. Every run has a recipe so you can reproduce it.
More from coding & agent
- AI gateways are becoming the key control layer in production AI stacks: 8 capabilities — goyalshaliniuk · 2026-11-21
- Mnemos 2.0: agent memory designed to preserve what matters for the next session — RileyRalmuto · 2026-11-21
- InferenceBench ranks agents on LLM serving speedup: Claude Opus 5.5 tops at 12.08x, agents lose to simple hyperparameter search — maksym_andr · 2026-11-21
- 4 architecture patterns to tame multi-agent system reliability, from a 27-year reliability engineer — bibryam · 2026-11-21
- AutoCut + FrameAtlas: open-source local-first video editing driven by Claude Code via MCP — moheshmohan · 2026-11-21
- InfoQ: What an agent harness is, and two ways to build one — bibryam · 2026-11-21