Local Qwen models can't finish a PacMan clone; maze design is the endless stumbling block
nixudos · reddit · 2026-09-27
Testing Qwen 27B 4KM and Qwen Next NVFP4 locally on the Deepseek harness, the task — "build a faithful browser PacMan clone within a limited token budget, delegating subtasks to sub-agents" — never completes regardless of instruction or compaction tweaks.
Qwen Next avoids loops, but maze layout design is always where it stalls: it rethinks the layout endlessly and never produces a final MD file. Context is limited (95K on the 27B, 65K on the Next). The author deliberately avoids hand-holding since this is a model test, and invites others to share configs (quant, KV size/quant, harness) if they got a Qwen to finish it.
More from coding & agent
- Adobe MAX: MCP across Creative Cloud, AI agents editing layers, Firefly Builder and Graph — umesh_ai · 2026-11-11
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11