Qwen 3.8 27B Fixes Bugs in 10 Minutes; Flash Next Takes 2.5 Hours
poofph · reddit · 2026-09-28
A newcomer running local models on dual RTX 5090s compared two Qwen models on the same bug-fixing task: Qwen3.8 27B finished in 5-10 minutes while Flash Next took 2.5 hours, despite 27B actually having higher throughput (2000-3000 infill, 100-150 tok/s output vs 1600-2300 and 60-100). Strangely, Flash Next wrote its code fixes about two hours before it finally stopped, apparently spending the remaining time running tests without further edits. The author, who isn't a programmer and uses AI-generated code, asks what causes this and what settings could match 27B's speed.
More from coding & agent
- Baseten's Guide: Evaluate Open Models for Coding Agents on Your Own PRs — baseten · 2026-11-16
- Google Research open-sources PaperOrchestra, a multi-agent framework for writing AI papers — google-research · 2026-11-16
- ffmpeg-skill gives coding agents local video editing via 42 FFmpeg tools and MCP — kajisho5 · 2026-11-16
- Agents Now Reply to Slack Messages on Their Own, Resolving a Prod Incident Ghost-to-Ghost — iamrobotbear · 2026-11-16
- Engineer's Interview Prep Method: Use Claude to Find Which Answers Undersell Your AI Work — brandon_galang · 2026-11-16
- DeepSeek open-sources its app frontend "chat-demo", already at 3.9k stars — NielsRogge · 2026-11-16