Zed unveils AI Edit Prediction feature powered by its new open-source language model, Zeta

Zed Editor has launched an innovative edit prediction feature aimed at boosting coding efficiency by predicting users' next changes. This feature is powered by Zeta, their new open-source language model derived from Qwen2.5-Coder-7B, and refined with an open dataset.

Users can quickly apply suggested changes by pressing the 'tab' key and can chain multiple edits with repeated presses. The feature integrates seamlessly with existing language server protocol (LSP) autocomplete functions, allowing users to hide the LSP menu and preview the predicted edit by holding the 'option' or 'alt' key.

Zeta's predictions are context-aware, enabling it to naturally rewrite larger code segments around the cursor. These predictions are validated by a larger language model to ensure alignment with various coding styles and the developer's intent. The model's training used Direct Preference Optimization (DPO) to identify effective editing patterns and minimize common errors. To maintain low latency, speculative decoding is used for faster predictions, supported by a global infrastructure with GPUs in North America and Europe. The edit prediction feature is available for free during its public beta phase, accessible by downloading Zed and signing in with a GitHub account.