HuggingFace released ColGrep-Agent-2B on October 12, 2026, saying it runs locally and reduces API token usage for code search. It is the company's first code-search model update since the release of ColGrep in 2025.
HuggingFace reported a 34% reduction in API tokens, measured on 52 code-search questions across 13 repositories. That compares with grep, which uses more tokens for the same tasks.
ColGrep-Agent-2B is built on the ColGrep semantic code search engine and targets code-search tasks in development workflows. Availability begins with a public release, initially for developers and researchers.
"We built ColGrep-Agent-2B to handle that part locally," said Raphael Sourty, lead engineer at HuggingFace. "It's a 2-billion-parameter model that searches a repository with ColGrep, reads promising files, follows references, and returns the relevant file paths and line ranges along with the code."
The announcement follows the release of ColGrep, which HuggingFace said was a key step toward improving code search efficiency. HuggingFace did not say when the model will be integrated into mainstream coding tools, and raised the open question of whether the model's performance will scale to end-to-end coding tasks.
Source: huggingface
