Zeta2.1: 3x Fewer Tokens, 50ms Faster

Zeta2.1, the latest update to Zed's edit prediction model, offers significant improvements in performance and efficiency compared to its predecessor, Zeta2. The new model emits significantly fewer output tokens, which enhances prediction speed by up to 50 milliseconds and reduces server usage by 30%. These advancements are attributed to a new prompt format called "Multi-Region," which focuses changes only around the code intended for edits, resulting in faster predictions with each keystroke. Zeta2.1 remains open-weight and is available for download on Hugging Face, having been trained on opt-in data from open-source repositories. The update also includes the publication of Rust code bindings to PyPI for easier self-hosting, making Zeta2.1 the default choice for Zed users, with options to explore Zed Pro or Zed Business for more extensive features. The model is designed to be run locally and is accessible on multiple operating systems, while Zed continues to seek passionate individuals to join their team in advancing software development.