Introducing Composer 2.5

Composer 2.5, now available in Cursor, represents a significant advancement over its predecessor, Composer 2, with improved capabilities in handling long-term tasks, following complex instructions, and collaborating effectively. The enhancements are achieved through scaled training, more sophisticated reinforcement learning environments, and new learning methods, focusing not only on intelligence but also on behavioral improvements such as communication style and effort calibration. Composer 2.5 is trained on more challenging tasks using 25 times more synthetic tasks, with innovations like targeted textual feedback to address specific mistakes during model training. Built on the open-source checkpoint Moonshot's Kimi K2.5 and in collaboration with SpaceXAI, the model benefits from a substantial increase in compute resources, boosting its capabilities significantly. The RL training incorporates synthetic data, with methods to prevent reward hacking through agentic monitoring tools, and utilizes technologies like Sharded Muon and dual mesh HSDP for pretraining optimizations. Composer 2.5 is priced variably depending on input and output token speed, offering a more cost-effective solution compared to other frontier models, and includes promotional usage incentives for new users.