1 tiếng nói · chưa gắn nhãn trưởng lab · 1 nền tảng · chưa có số tương tác · Nguồn crawl
A
aartiles24 0 bài viết · X Nguyên văn nguồn Chi tiết nguồn
The Pi-adapted Qwen 3.8 27B actually works. Here’s the proof.
Lots of people have been saying the harness-hub Pi adapter for Qwen 3.8 27B is great. Most of what I found was benchmark scores and tok/s, with little about the actual tasks behind them. So I gave it a real task I had already run with the base 27B on my MacBook (M5 Pro, 64 GB): a full raycaster FPS from one brief (prompt by @ddalcu, shared by @Beamsters1), built by Tenaz + Pi with no human in the loop.
The base 27B, 4 attempts:
❌ 2 blocked before writing a single line of code
❌ 2 stopped after ~4 h with 2 tasks done and no playable game
It kept burning its whole output budget on reasoning.
The pi-adapted 27B (OptiQ 4-bit on Splash, thinking off):
✅ COMPLETE in 2 h 3 min: 34 sessions, 0 failures, 0 human input
🎮 All 8 must-have features work when you play it
⚡ ~52 tok/s, 0.38 s median TTFT, 96% prompt cache
🧠 Memory pressure normal the whole time, with Chrome and Claude open
Same brief, same laptop (different engine and harness too, details in the link): from no playable game to a finished one.
The adapter does what it claims: short turns, no thinking, real tool use. Qwen 3.8 Flash Next is still the smarter model, but it needed ~7 h for this game. The adapted 27B got there in about 2, and left the Mac usable.
That makes it the local model I’d hand to a frontier model as its worker.
Every attempt, builds and numbers side by side:
https://aartiles.github.io/local-model-benchmarks-site/challenges/raycaster-shooter/
Media