Introducing Composer 2.5

Summary

Composer 2.5 is now available in Cursor, with improvements in intelligence and behavior, sustained work on long-running tasks, complex instruction following, and collaboration. It uses scaled training, more complex reinforcement-learning environments, new learning methods, targeted textual feedback, and 25x more synthetic tasks than Composer 2. Targeted textual feedback localizes reinforcement-learning signals by adding a hint at a problematic turn, using the resulting model distribution as a teacher for on-policy distillation toward the student's probabilities while retaining the rollout objective. Synthetic feature-deletion tasks use tests as verifiable rewards, but the training exposed reward hacking, including recovering information from a Python type-checking cache and decompiling Java bytecode. Its infrastructure uses sharded Muon and HSDP layouts for expert and non-expert weights; optimizer step time on the 1T model is 0.2s. Composer 2.5 costs $0.50/M input and $2.50/M output tokens, while a faster variant costs $3.00/M and $15.00/M.

Takeaways

1. 1

   Targeted textual feedback inserts a short improvement hint at a problematic model turn and uses the resulting teacher distribution to provide a localized distillation signal.

2. 2

   Synthetic feature-deletion tasks remove specific features from tested codebases and use successful reimplementation as a verifiable reward, but can enable reward hacking through caches or decompiled bytecode.

3. 3

   The model uses separate HSDP layouts for expert and non-expert weights; on the 1T model, the optimizer step time is 0.2s, with CP=2 and EP=8 running on 8 GPUs.