ColGrep-Agent-2B cuts Claude Code API tokens 34% and costs 40% with local code search
IgorCarron · x · 2026-10-12
maxencelsb introduced ColGrep-Agent-2B, an open 2B model integrated into NextPlaid via ColGrep that runs locally to make coding agents more efficient.
- Delegating code search to the tiny agent yields 34% fewer API tokens and 40% lower API costs in Claude Code
- Architecture: a 2.5 GB GGUF model on llama.cpp repeatedly searches and reads a local index, returning locations + code to the cloud frontier LLM — no code leaves the machine
- Benchmarks show near-27B/31B retrieval F1 on SWE-bench Lite and Multi-SWE-bench Flash despite 2B parameters
More from coding & agent
- Indie builder ships infinite-canvas book recs site made with Cursor, Next.js and Supabase — floguo · 2026-10-12
- Manus 2.0 builds and publishes a playable 3D open-world game from a single prompt — minchoi · 2026-10-12
- Weaviate Indexes Metadata Natively: Filtered Vector Search Without a Separate Relational DB — CShorten30 · 2026-10-12
- Watching a Personal Agent 'gob' Book Japan Train Tickets End-to-End in a Cloud Browser — gabrielchua · 2026-10-12
- Compound Engineering v3.30.5 adds auto-retiring learnings and AI slop test detection — kieranklaassen · 2026-10-12
- DHH's RailsWorld AI keynote sparks community backlash: 'Weimar era of Ruby' — kieranklaassen · 2026-10-12