Edit Prediction
Zed predicts your next edit based on your activity. Powered by Zeta2, our open-weight model, or any provider you choose.
What developers are saying
Predictions that feel right
Here's what developers love about edit predictions.
So far the Edit Prediction feature has been great. Big fan of the privacy-focused default & controls around sharing training data, especially for non-open-source projects. Predictions are fast + have been contextual.
Matt Silverlock
Sr. Director of Product
The inline completions are pretty magical. I'm converting/updating a project to 11ty, and it's doing better at contextual predictions of what I want to add to files better than any other AI assistive tool I've used so far.
Bob Rudis
V.P. Research
Zed's Edit Prediction is impressively accurate and feels super snappy. Where other autocomplete solutions often struggle with reliability, Zed's implementation stands out.
Andreas Thomas
CTO & Cofounder
Predicting intent, preserving focus
Edit predictions that fit into your existing workflow. Other tools complete code. Zed predicts what you're trying to do next.
Plow through edits by repeatedly hitting tab
As you work, Zed now predicts your next edit, so you can apply it just by hitting tab. Once you accept a prediction, you can perform multiple follow-up edits by pressing tab repeatedly, saving you time and keystrokes.
Thoughtful integration with Language Server completions
When language server completions are visible, Zed won't preview the predicted edit until you press option or alt. As soon as you press the modifier, Zed previews the edit and hides the menu to enable an unobstructed review.
Subtle Mode: predictions on demand
If you find predictions distracting, Subtle Mode keeps them hidden until you ask. Hold alt-tab to see the prediction, release to dismiss. The model keeps working; you just choose when to look.
How Zeta2 works
Built for edit prediction, trained on real edits
Zeta2 is the only model purpose-built for how Zed works, trained entirely on open source code with public weights.
View on Hugging FacePredictions speak your language
Zeta2 uses the language server to retrieve the type and symbol definitions around your cursor, so predictions understand your code's actual structure rather than guessing.
See Pull Request →Evaluation built into the pipeline
We score predictions only on the code that changed, using diff-aware metrics and line-level exact match to measure real accuracy rather than copy fidelity.
See Pull Request →Training data from real edits
Multi-file commits are split into single-change examples, stratified by repository, and supplemented with distillation from a larger teacher model.
See Pull Request →Experimenting with different, more efficient prompt formats
Allowing the model to output a subset of the editable region to reduce the number of tokens it has to generate
See Pull Request →So light, so fast, so helpful...
Pleasantly type together with Edit Predictions on Zed, right now.
Hit Tab to Accept
