Mercury Edit 2

The tool predicts the developer’s next code change by analyzing recent edits and the surrounding codebase context. It generates edit suggestions in parallel using a diffusion‑based language model, allowing the output to appear instantly and be accepted with a simple keystroke. The system is trained on a curated dataset of multi‑language edit examples and further refined with human preference feedback through an unpaired reinforcement learning method, which reduces overly frequent or lengthy suggestions.

Target users are programmers who need rapid, context‑aware assistance during interactive development, especially in latency‑sensitive parts of the workflow such as line completion, variable renaming, refactoring, or feature implementation. By delivering suggestions that are both high‑quality and low‑latency, the model aims to integrate seamlessly into a developer’s thought process without causing distraction.

Distinctive aspects include the use of diffusion to produce multiple tokens concurrently, resulting in faster response times than traditional auto‑regressive models, and a training pipeline that incorporates explicit accept/reject feedback to align outputs with human preferences. Benchmarks report a 48 % higher acceptance rate and 27 % greater selectivity compared with the previous version, while also outperforming alternative next‑edit models in speed and accuracy.