Local 27B coder hits compaction loops on big projects; code indexing proposed as fix
OuterKey · reddit · 2026-10-14
A Reddit user running Qwen 3.8 27B locally on an R9700 for coding describes a compaction death spiral: as projects grow, context fills faster, compaction fires between every edit until the model stops doing useful work.
Switching to Q4KXL quantization and Q8 KV cache extended context from 100-150k to 200k tokens. Proposed mitigations include indexing code (function signatures with comments, line numbers, full structs) to reduce re-reads, and extreme modularity to narrow model focus — though cross-module changes make the latter cumbersome.
More from coding & agent
- OpenAI Cookbook adds 25 TypeScript examples: one-line short films, talking photos, video ads — Daniel_Farinax · 2026-10-15
- AIPOCH Open-Science launches as open-source local-first AI research workbench with reproducible .science packages — JaynitMakwana · 2026-10-15
- Anthropic publishes official guide for building offline AI agents with a Slack+GitHub briefing template — msharmas · 2026-10-15
- 3D light show of 10,000 simulated drones over harbor city built with Claude Opus — techartist_ · 2026-10-15
- Coding agent manager Orchestrator hits 13,000 GitHub stars in 8 months — julianweisser · 2026-10-15
- awesome-agent-apis: open catalog of 660+ generative-media model APIs for agents — matchaman11 · 2026-10-15