Blog / Research Composer is an advanced agent model developed for software engineering, achieving significant improvements in both intelligence and speed. This model excels in completing real-world software engineering tasks within expansive codebases, boasting four times faster generation speeds than comparable models. Composer is optimized for interactive use in Cursor, building on insights from the prototype model "Cheetah" to enhance developer workflows by ensuring an engaging coding experience. As a mixture-of-experts language model, Composer leverages reinforcement learning across varied development environments, providing solutions such as code edits, plans, or informative responses. The training process incorporates a wide array of tools and focuses on maximizing efficiency and parallelism to maintain interactive development fluidity. Composer's development involved significant investment in infrastructure, enabling scalable training with thousands of GPUs and enhancing inference speeds. The model's capabilities are tested through the Cursor Bench, which evaluates its usefulness to developers, including its ability to adhere to codebase abstractions and practices. Composer's release aims to provide a valuable tool for software engineering, a sentiment echoed by its adoption among the developers at Cursor.