The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Trends

Supporting case

A small local agent can take repository search out of a frontier coding model’s workflow

What happened

LightOn released ColGrep-Agent-2B on October 12. It searches a software repository, reads relevant files and returns filenames and line ranges to a developer or larger coding assistant. The available model and installation documentation ↗ provide an Apache 2.0 license, local inference, Windows/macOS/Linux installation routes and a default compressed model of about 2.5 gigabytes. It does not require an API key or separate inference server for that local search component.

LightOn tested 52 questions across 13 repositories, three runs per question, using Claude Code with different search arrangements. It reports 34% fewer billed API tokens and 40% lower cache-adjusted API cost when search was delegated, without a statistically significant accuracy change. Questions and answer grading involved larger AI models; the exercise is not an independently audited field study of software maintenance. Local inference work was not counted as billed API tokens. The main coding agent can still depend on a hosted provider.

For The Race, the concrete separation is one usable workflow component: locating code under local control. It does not establish independent coding capability or movement of an entire organization between platforms.

Dates and implementation stage

Model/component released October 12, 2026

Supporting sources

Related trends