Mercury Edit 2 is a purpose-built diffusion LLM designed to serve as a high-speed assistant for the most latency-sensitive aspects of software development. By utilizing diffusion to generate tokens in parallel, the model delivers next-edit predictions that integrate seamlessly into a developer's workflow, allowing for immediate acceptance of suggestions. It functions by analyzing a user's recent edits alongside the broader codebase context to anticipate and propose the next logical change, effectively acting as a responsive partner that keeps pace with the speed of human thought.

The model was developed using a carefully curated dataset of edits across diverse programming languages and scenarios to ensure high-quality, relevant suggestions. To refine its behavior and prevent overzealous or distracting edits, the training process incorporates a human preference dataset that tracks explicit feedback on accepted versus rejected suggestions. This data is leveraged through an unpaired reinforcement learning method known as KTO, which aligns the model's output with user preferences. This approach ensures that the model remains a practical, unobtrusive tool for developers, balancing predictive power with the precision required for real-world coding environments.