Cursor releases Composer 2 technical report on agentic coding model training

Cursor has published a technical report detailing the training of Composer 2, an agentic software engineering coding model built upon the Kimi K2.5 base.

• The model undergoes continued pretraining to deepen coding knowledge followed by large-scale reinforcement learning in realistic Cursor sessions.
• Evaluation uses CursorBench, derived from real engineering team sessions with terse prompts and multi-file changes.
• Composer 2 scores 61.3 on CursorBench, a 37% improvement over Composer 1.5, and achieves competitive results on SWE-bench Multilingual and Terminal-Bench.
• Training infrastructure includes custom low-precision kernels for MoE training on Blackwell GPUs and the Anyrun platform for sandboxed environments.

The report highlights that Composer 2 offers a Pareto-optimal tradeoff between accuracy and inference cost for interactive developer workflows.