Zed ships Zeta2 for code editors
Zed released Zeta2, a distilled model for its code editor that the company says has 30% better acceptance rates than its prior model by focusing on better training rather than larger size. The release underlines a trend toward targeted, efficiency‑focused models for developer tooling. (x.com)
Most code helpers guess the next few characters. Zed built Zeta for a narrower job: predicting the next edit, which can mean inserting, deleting, or reshaping multiple lines inside the file you are already changing. (zed.dev) Zed says its new Zeta2 model improves acceptance rate by 30% over Zeta1, and it made Zeta2 the default edit prediction model for all Zed users on March 25, 2026. The company says the model is also faster, which matters because these suggestions only feel useful when they appear before your hands move on. (zed.dev 1) (zed.dev 2) An acceptance rate is the share of suggestions people actually take by pressing Tab. It is a blunt but practical scorecard: if the editor keeps offering text you ignore, the model is wasting your time instead of saving it. (zed.dev) The surprising part is that Zed did not say it won by making Zeta2 bigger. In its March 25 post, the company said the core architecture stayed the same and the gains came from rebuilding the training pipeline and improving the data. (zed.dev) That is a different bet from the usual “more parameters, more compute” race. It is closer to tuning a race car for one track, because edit prediction has a very specific job: watch your recent changes, look at nearby code, and suggest the exact patch you were about to type. (zed.dev 1) (zed.dev 2) Zed’s first Zeta release in February 2025 was already built around that idea. The company described it as an open-source, open-data model for edit prediction rather than a general chatbot living inside the editor. (zed.dev) (huggingface.co) By February 2026, Zed had also opened the door to competing providers inside the editor, including Mercury Coder, Sweep, Ollama, Codestral, and GitHub Copilot’s Next Edit Suggestions. That gave users a direct side-by-side test of whether a specialized in-house model could beat bigger outside systems on the one thing Zed cared about. (zed.dev) Zed’s April 7 write-up says building Zeta2 required two ingredients: infrastructure that lets the team learn quickly and a very large training set. In other words, the company is arguing that for coding tools, better examples and faster iteration can move the needle more than simply scaling model size. (zed.dev) That fits a wider shift in developer tools toward smaller models aimed at one repeated action inside a workflow. If the task is “predict the next edit in this file” instead of “answer anything,” a model can be judged on latency, hit rate, and whether it stays out of the way. (zed.dev 1) (zed.dev 2) So the Zeta2 launch is not just one editor shipping one new model. It is a live example of software companies trying to win with narrower, cheaper, faster models that do one job well enough that developers actually keep the feature turned on. (zed.dev) (zed.dev)
