ColGrep-Agent-2B delegates code search to a small local model

12 October 2026LightOnCode search

LightOn proposes an intermediary that searches the repository and returns files, lines and code to the main agent.

Open model published

The two-billion-parameter model uses ColGrep, a multi-vector semantic search tool, reads promising files and follows references.
Across 52 questions in thirteen repositories unseen in training, LightOn reports 34% fewer API tokens and 40% lower API cost versus grep, accounting for caching.
The post provides a model, GGUF variant, code and paper; no statistically significant accuracy difference is reported under this protocol.

Repository search becomes a delegable task with its own budget.

Measure recall, latency and total cost against grep and semantic search alone.

Local execution and public artifacts are offered; check license, hardware resources and configuration before adoption.

The test is limited and published by the authors. API savings imply neither faster completion nor zero total cost for the local model.