Model overview
zeta-2 is a code edit prediction model developed by Zed Industries that suggests how to rewrite code in a specific region. The model builds on Seed-Coder-8B-Base through fine-tuning to understand code context, edit history, and cursor position. This capability differs from the earlier zeta model, which was fine-tuned from Qwen2.5-Coder-7B. Both models focus on the same core problem: helping developers complete edits rather than generating code from scratch.
Model inputs and outputs
zeta-2 accepts structured input that combines multiple code context signals. The model uses a suffix-prefix-middle (SPM) style prompt with git-merge-style markers to indicate the editable region. The input includes the code before the cursor, the code after the cursor, related file content, and recent edit history to inform the prediction.
Inputs
- Code before the editable region with cursor position marked explicitly
- Code after the editable region (suffix context)
- Related file content to provide broader codebase context
- Edit history in git diff format showing recent changes
Outputs
- Rewritten content for the editable region that replaces the original code
- Revised code formatted with git-merge-style markers indicating the updated section
Capabilities
The model predicts how developers should revise code based on context and recent edits. It examines the surrounding code, related files, and past modifications to suggest appropriate rewrites. This allows it to handle refactoring, bug fixes, and style improvements by understanding the developer's intent from their recent changes and the broader code structure.
What can I use it for?
Use zeta-2 to power code editing features in development tools. Integrated into an IDE or editor, it can suggest how to rewrite selected code regions based on recent edits and project context. This works for refactoring code, applying fixes across similar code patterns, updating variable names consistently, or adapting code style to match project conventions. Developers can monetize this through editor extensions, specialized IDEs, or AI-assisted development platforms that charge for advanced coding suggestions.
Things to try
Test the model with code regions that require context-aware rewrites. Provide recent edit history to see how it learns patterns from your changes—for instance, if you recently renamed a variable in one function, provide that as edit history when asking it to update similar code elsewhere. Try it on different programming languages and file types within your project to understand how it generalizes across contexts. Experiment with varying amounts of related file context to find the balance between suggestion quality and performance.
This is a simplified guide to an AI model called zeta-2 maintained by zed-industries. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.