Federico Cassano

How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL

Federico Cassano, Dmytro Dzhulgakov, Sonya Huang

Cursor has pivoted from a pure application company to a foundation model developer by training Composer 2, a specialized software engineering model built on a Kimi 2.5 base using a global, asynchronous pipeline to maximize compute efficiency. The team deployed a custom infrastructure that integrates Reinforcement Learning from real-time user feedback and simulated rollouts while overcoming synchronization challenges through lossless delta compression and custom GPU kernels. This strategic approach allows Cursor to saturate model capacity with high-value coding data, achieving competitive performance at a fraction of the cost of general-purpose models while moving the industry toward baked-in specialized behaviors rather than prompt engineering.