Local Qwen 27B reads 26K-line design doc, builds 3D game in 12 hours
Healthy-Nebula-3603 · reddit · 2026-09-09
Inspired by Bijan Bowen's Subway FPS video, a Reddit user pushed locally-run Qwen 3.8 27B (Q4KXL via llama-server) to its 3D game-development limits.
The setup: first generate a 267 KB DESIGN.md (26K lines) with Fable 5.1, then feed it to Qwen running in a PI agent with 120K context + vision, CPU inference with MTP for speed. Over 12 hours the model read 11M tokens and wrote 3.2M tokens, producing a playable 3D game. The author shared the design doc on Google Drive for reproducibility, calling it insane what's possible on a home computer.
Related event: Local Qwen 27B Builds 3D Game in 12 Hours, Reading 11M Tokens(2 posts)→
More from coding & agent
- He benchmarks new models by having them build their own copilot — $28 of compute — bucolucas · 2026-10-27
- Founder generates 46 app promo videos in 4 days with Claude, skipping $11,500 in UGC fees — PrajwalTomar_ · 2026-10-27
- Should agents be allowed to change plans mid-task? Reddit debates the drift boundary — wistfulcracker84 · 2026-10-27
- HeyGen + Canva MCP turns slide speaker notes into an avatar presentation video — HeyGen · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- Cloudflare Computer adds agent interruption recovery and approval pauses in v0.7.0 — irvinebroque · 2026-10-27