A technical report on Composer 2

Summary

Composer 2 is a coding model for agentic software engineering, and its technical report describes training from continued pretraining of the open Kimi K2.5 base model through large-scale reinforcement learning. The two-phase recipe emphasizes code during pretraining, then runs RL in realistic Cursor sessions using the deployed model’s tools and harness across a broad developer-task distribution. CursorBench, built from real engineering-team sessions, evaluates terse or ambiguous prompts requiring changes across many files, addressing limitations of small, narrowly specified public tasks. Composer 2 scores 61.3 on CursorBench, 37% above Composer 1.5, plus 73.7 on SWE-bench Multilingual and 61.7 on Terminal-Bench, at significantly lower inference cost than comparable models. The report also details Blackwell-optimized low-precision MoE kernels, asynchronous multi-region RL, and Anyrun’s sandbox infrastructure.

Takeaways

1. 1

   Reducing pretraining loss improved downstream RL performance, with stronger base knowledge reliably translating into a better agent.

2. 2

   CursorBench tasks come from engineering-team coding sessions and include terse, ambiguous prompts requiring hundreds of lines of changes across many files.

3. 3

   Composer 2 scored 61.3 on CursorBench, a 37% improvement over Composer 1.5, while scoring 73.7 on SWE-bench Multilingual and 61.7 on Terminal-Bench at significantly lower inference cost than comparable models.