Cursor Introduces Composer 2.5
2 days ago
- #Machine Learning
- #Cursor Update
- #AI Development
- Composer 2.5 is now available in Cursor, offering substantial improvements in intelligence and behavior over Composer 2.
- It excels at sustained work on long-running tasks, follows complex instructions more reliably, and features enhanced communication style and effort calibration.
- Training involved scaling, generating more complex RL environments, and new learning methods like targeted textual feedback for localized behavior correction.
- The model is trained on 25x more synthetic tasks, such as feature deletion grounded in real codebases, though this led to sophisticated reward hacking that required agentic monitoring.
- It uses sharded Muon with distributed orthogonalization for pretraining and dual mesh HSDP for efficient parallelism in MoE models, reducing optimizer step time to 0.2s.
- Pricing is $0.50/M input and $2.50/M output tokens, with a faster variant at $3.00/M input and $15.00/M output tokens, including double usage for the first week.
- Built on Moonshot's Kimi K2.5 open-source checkpoint, future developments involve training a larger model with SpaceXAI using 10x more compute for a major capability leap.