Coding agents burn a lot of tokens just figuring out where to look. LightOn’s answer is ColGrep-Agent-2B, an open 2-billion-parameter code-search agent that runs locally between your coding agent and the repo.
How it works
Ask it a question and it searches with ColGrep (LightOn’s multi-vector semantic code search), reads promising files, follows references, and hands back file paths, line ranges and the code itself. Your frontier model gets the answer instead of doing ten rounds of grep.
The results
Plugged into Claude Code on 52 code-search questions across 13 repos it hadn’t seen in training, delegating search cut API tokens by 34% and cache-aware cost by 40% versus grep, with no statistically significant change in accuracy.
My take
This is the pattern I like: let a small local model do the grunt work and save the expensive model for thinking. Token usage is something I watch closely, so a third off the top is worth a look.
