Zeta2.1: 3x Fewer Tokens, 50ms Faster (zed.dev)

Zed has announced the launch of Zeta2.1, an upgraded version of its edit prediction model, which significantly enhances performance by emitting three times fewer output tokens and reducing response times by up to 50 milliseconds. This efficiency leap, achieved through a new "Multi-Region" prompt format, allows the model to focus only on the specific code changes instead of processing a larger surrounding area. These changes enable Zeta2.1 to operate with 30% fewer servers while maintaining a higher acceptance rate and reducing explicit rejection rates. The roll-out of Zeta2.1, which is open-weight and trained on opted-in data from open-source repositories, promises a seamless experience for local users. It works out of the box and comes with Rust code bindings released to PyPI, making self-hosting even more accessible. As the default edit prediction model for Zed, Zeta2.1 not only improves prediction speed and resource utilization but also invites community engagement through data sharing, paving the way for ongoing enhancements in AI-driven software development tools.