Multiple Edit Prediction Providers Now Available

Zed has introduced a pluggable edit prediction provider architecture, giving developers the ability to choose between multiple code prediction services. The supported providers are:

• Zeta (default) — Zed's proprietary model trained specifically for edit prediction
• Mercury Coder — Uses diffusion-based architecture for parallel output refinement
• Sweep — Fast predictions trained with online RL, delivering suggestions in under 100ms
• Ollama — Run open-weight models locally (Qwen, CodeLlama, DeepSeek) without sending code off-machine
• GitHub Copilot — With Next Edit Suggestions
• Codestral — Mistral's code model

Simplified Provider Architecture

The refactoring moved most edit prediction logic (state management, UI integration, debouncing, caching) into Zed's core, significantly reducing the effort required to add new providers. Developers implementing new providers only need to handle model-specific tasks: prompt construction, API calls, and response parsing. This opens the door for community contributions via pull requests.

Key Capabilities and Trade-offs

Edit predictions work best when they match individual coding patterns—anticipating edits, suggesting closing braces, or removing unnecessary code. Each provider offers different latency characteristics and performance depending on language, project size, and editing style. Mercury Coder's 1-month free trial is available with a provided API key.

Getting Started

Developers can configure their preferred provider directly in Zed's Settings Editor and experiment to find the best fit for their workflow. The local Ollama option is particularly valuable for teams with code privacy requirements.