ColGrep-Agent-2B

ColGrep-Agent-2B

A 2.5B-parameter model that finds the files and line ranges relevant to a bug report or question about a code repository, by searching and reading the repo. Runs locally through ColGrep.

Made by lightonai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: ColGrep-Agent-2B is a small code-localization agent from LightOn. Given a bug report, a feature request or a question about a repository, it finds the files and line ranges that matter.

It alternates between searching with ColGrep and opening files in a terminal that cannot change anything, then reports the precise spots. It is based on MiniCPM5-2B and is meant to take repository exploration off larger coding agents.

Who it suits: Developers and coding-agent builders who want to find the right code quickly, and who prefer a small local model over sending a whole repository to a big one.

Industry | Who uses it | How they use it | Example
Software | Developers | Locate code for a bug or question | Finding where retries are configured
AI tooling | Coding-agent builders | Hand repository search to a small helper | A larger agent that delegates code lookup
Engineering teams | Maintainers | Triage issues faster | Pointing an issue at the likely files

You can skip it if you do not work with code repositories, or if you want a model that writes code rather than finds it.

Where to use it: It runs locally through the ColGrep command line on macOS, Linux and Windows, with GPU acceleration when available. It needs no Python environment, inference server or API key.

• The colgrep command line
• macOS and Linux with Homebrew, or Windows with PowerShell
• Cargo, to build ColGrep from source
• Larger coding agents that call colgrep --agent

Use case | Who | What they get back
Issue localization | Maintainers | File paths and line ranges for a bug report
Question answering about code | Developers | Locations in the repo that answer the question
Delegated search | Agent builders | Compact results for a bigger model to use

How to set it up: You install ColGrep once and let it fetch the model. The install-agent step downloads a Q8_0 GGUF file of about 2.5 GB, checks that it loads and prepares a prompt cache.

• Homebrew, or Cargo, or the card's installer scripts
• About 2.5 GB for the default model file
• A GPU is optional but speeds things up

1. Install ColGrep with Homebrew or Cargo, using one of the commands below.
2. Run the install-agent step, which the same commands include.
3. Index your repository once with colgrep init.

brew install lightonai/tap/colgrep && colgrep --install-agent

cargo install colgrep && colgrep --install-agent

cd /path/to/project
colgrep init

How to use it: Ask a question in plain language with colgrep --agent. After the first index, searches notice changed files and update the index before answering. Results come as path, start line and end line.

colgrep --agent "where does the agent decide which GPU backend llama.cpp uses"

Tool | What it does
colgrep --install-agent | Downloads and prepares the default model
colgrep init | Indexes the current repository
colgrep --agent | Runs the agent on a question and returns code locations

Example: On ColGrep's own code, the card asks where the agent decides which GPU backend llama.cpp uses. In the shown session on an M3 Pro, the agent made 9 turns and 2 searches and returned 5 locations in 21.7 seconds.

For newcomers and experts: If you are new, install ColGrep, index a small project and ask one concrete question. The answer is a list of files and line numbers you can open yourself.

If you know the field, the agent was trained with reinforcement learning to search with ColGrep and read files through a read-only terminal. It loops over searches and reads, then calls finish with the final locations, which fits a larger agent's tool loop.

Index the repository with colgrep init
Ask a question with colgrep --agent
Agent searches and reads files
Agent finishes with path and line ranges

Good to know: The output follows a path, start line and end line format, so scripts and other agents can use it directly. The card shows tool results omitted for readability in its sample session.

• License: Apache 2.0
• Cost to run: none named, and no API key is needed
• Data it reaches: the repository you index and ask about, plus the model file it downloads
• Not for: writing code, since it only locates relevant code

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 2.5B
• Weights: about 5 GB in BF16; about 1.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,141
• Likes: 38
• License: apache-2.0
• Languages: en
• Based on: openbmb/MiniCPM5-2B
• Quantized versions: 6
• Tags: transformers, safetensors, llama, text-generation, code, code-search, tool-use, reinforcement-learning, colgrep, conversational

Links

• Model: https://huggingface.co/lightonai/colgrep-agent-2B
• Files: https://huggingface.co/lightonai/colgrep-agent-2B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lightonai/colgrep-agent-2B

Then follow the usage in its model card. Nothing in this kit runs by itself.