LightOn AI has introduced ColGrep-Agent-2B, a locally running 2-billion-parameter model that sits between a coding agent and a code repository, using a multi-vector semantic code search tool called ColGrep to locate relevant files, follow references, and return file paths and code. In tests with Claude Code across 52 code-search questions, delegating search to the agent cut API token usage by 34% compared with standard grep, with no statistically significant change in accuracy. The approach shifts costly, multi-turn search work from a frontier model to a small local one, though end-to-end response time rose from 13.0 to 21.6 seconds.
