CONFIRMED: LightOn announced the release on its Hugging Face blog on 12 October 2026, and the model is live on Hugging Face. All performance figures below are LightOn’s own, self-reported results and have not been independently tested. They are for code-search questions, not full coding tasks.
LightOn has released ColGrep-Agent-2B, “an open 2-billion-parameter code-search agent that runs locally and sits between your coding agent and the repository,” according to its announcement [1]. The model is published under the Apache-2.0 licence [2].
A code-search agent is a small AI model whose only job is to find the right place in a codebase. Instead of a big, paid cloud model such as Claude hunting through files itself, it hands the question to this local model, which comes back with file paths, line ranges and the relevant code.
What it does
Before a coding agent can fix a bug, it often spends several turns working out where to look: searching for a name, opening a file, following a reference, trying again. Each of those steps goes through the expensive frontier model, and the growing conversation is sent back on every turn, LightOn explains [1].
ColGrep-Agent-2B takes over that search step on your own machine. It uses ColGrep, LightOn’s semantic code search tool, which can match a description (“custom attributes”) to code that uses different words (“fieldDefinition”). The agent has three tools, colgrep for search, a read-only terminal, and finish to return results, and a 10-turn budget [1].
LightOn says there is no Python environment to set up and no external inference server or API key. The default model file is about 2.5 GB and runs on a GPU where available, or on the CPU [1].
What LightOn says it found (company claim)
LightOn ran Claude Code with Sonnet 5.5 on 52 code-search questions across 13 public repositories created in 2026. Each question ran three times in each of four setups, 624 sessions in total [1].
LightOn reports that handing search to the local agent “cut API tokens by 34% and cache-aware API cost by 40% relative to grep, with no statistically significant change in accuracy” [1].
From its results table [1]:
| Per question | Grep (built-in tools) | ColGrep-Agent |
|---|---|---|
| Tokens | 142k | 93k |
| Cost (cache-aware) | 7.3¢ | 4.4¢ |
| Correct answers | 83.3% | 80.8% |
| Time | 13.0 s | 21.6 s |
The accuracy gap is −2.6 percentage points, with a 95% confidence interval of −10.9 to +5.1. LightOn notes the difference “isn’t statistically significant, but the interval is wide enough that we can’t rule out a meaningful drop” [1].
A “Balanced” mode, which lets Claude choose between plain ColGrep and the agent, saved 21.2% of tokens while matching grep’s 83.3% accuracy, LightOn says [1].
On a code-localisation version of SWE-bench Lite, where the task is to find the relevant files and lines rather than write a fix, LightOn reports a File F1 of 76.18 and a Line F1 of 19.92 using its own ColGrep harness [1][2].
LightOn is explicit that “these results are for code-search questions, not end-to-end coding tasks” [1].
How to get it
LightOn lists Homebrew and Cargo installs, among others [1][2]:
brew install lightonai/tap/colgrep && colgrep --install-agentcargo install colgrep && colgrep --install-agent
It also ships hooks to connect the agent to Claude Code, OpenAI’s Codex and OpenCode (colgrep --install-claude-code, --install-codex, --install-opencode). For Claude Code, the hooks suggest the search tool rather than forcing its use [1].
How it was built
LightOn started from the open MiniCPM5-2B model, trained it first with a short supervised fine-tuning (SFT) “cold start”, then with reinforcement learning using a variant of GRPO, rewarding it for finding the right files and lines while penalising unnecessary steps [1][2].
What this means
- Cheaper searching, not faster coding. LightOn’s claim is about the cost of finding code. It does not claim any speed-up in writing or fixing code, and its own figures show each question took longer (21.6 s vs 13.0 s).
- A small model doing one job. The pitch is that not every step needs a frontier model: a specialised 2B model on your machine can do the searching, and your code does not have to leave the machine for retrieval or agent inference, LightOn says [1].
- Watch the accuracy caveat. The token savings are clear in LightOn’s tests, but the accuracy result is uncertain. The Balanced mode is the company’s own lower-risk option.
- Not yet independently checked. All results come from LightOn’s own benchmark.
Sources
- LightOn (Maxence Lasbordes, Raphael Sourty, Aarush Sinha, Amélie Chatelain), “ColGrep-Agent: Cutting Frontier Coding Agents’ Tokens by a Third”, Hugging Face blog, 12 October 2026. https://huggingface.co/blog/lightonai/colgrep-agent
- LightOn, lightonai/colgrep-agent-2B model card, Hugging Face. https://huggingface.co/lightonai/colgrep-agent-2B
