LightOn releases local ColGrep-Agent-2B for coding-agent repository search

LightOn has released ColGrep-Agent-2B, a small local model intended to carry out repository navigation for larger coding assistants. Rather than asking a frontier model to repeatedly search and open files, a coding assistant can delegate a description of the task to a separate search component. The local 2B agent uses LightOn’s multi-vector ColGrep search, inspects promising source files and follows references before returning file paths, line ranges and excerpts. The authors provide installation and integration instructions, a model link and a GGUF build.

LightOn tested the arrangement with Claude Code and Sonnet 5.5 on 52 code-location questions across 13 public repositories. Each question was repeated three times under four search-tool configurations, yielding 624 sessions. Relative to Claude Code using its normal grep, find and file-read tools, the authors report 34.3% fewer API tokens when search was delegated to ColGrep-Agent. They also report lower cache-aware API spending. These numbers are LightOn’s own evaluation, not independently reproduced Open Model Weights measurements.

The detailed results expose important limits. According to LightOn, mean answer correctness was 80.8% with ColGrep-Agent compared with 83.3% for the grep baseline, a difference the authors describe as statistically inconclusive. Average wall-clock question time rose from 13.0 seconds to 21.6 seconds. The evaluation measures finding code locations, not autonomous patching, successful tests or resolution of complete software-engineering issues. Token savings therefore should not be repackaged as a universal coding-agent performance gain.

The release illustrates a useful architecture for open-weight agent systems: a lightweight specialist can handle repetitive retrieval locally while a larger model remains responsible for reasoning or code changes. LightOn documents a read-only tool environment and examples for connecting ColGrep to existing assistants. Teams considering deployment must still inspect the exact checkpoint files and license, pin the runtime version, test their own repositories and account for additional local latency. Open Model Weights treats this as a source-backed model-intake candidate until those artifact-level fields are verified.