Mercury Edit 2

Mercury Edit 2 dramatically accelerates coding workflows with its next-edit prediction capabilities. Key features: • Predicts next code changes in real-time • Analyzes recent edits and codebase context • High acceptance rate for suggested edits • Exceptionally low latency for predictions • Diffusion-based generation for parallel token output This tool upgrades an existing next-edit model, complementing auto-complete functionalities by focusing on the most latency-sensitive parts of development. It curates high-quality edit datasets across multiple languages and scenarios, learning to interpret edit histories and codebase context to provide targeted suggestions. Through an unpaired reinforcement learning method (KTO) aligned with human preferences, the model's suggestions are highly relevant and less distracting. This approach significantly increases the utility of the model, leading to a 48% higher acceptance rate and 27% more selective, on-point edits. Built for developers and engineering teams, Mercury Edit 2 integrates seamlessly into your coding environment, allowing you to accept predictions with a simple Tab. It's ideal for those seeking to enhance coding speed and maintain focus by receiving timely, accurate code modification guidance.