Composer: Building a fast frontier model with RL

Summary

Composer is a mixture-of-experts agent model for software engineering that targets interactive use, with reported frontier coding results and generation speed four times faster than similar models on the authors’ benchmarks. It is trained with reinforcement learning on real-world software engineering challenges in large codebases, using production search and editing tools, terminal commands, and codebase-wide semantic search. Training responses may be code edits, plans, or informative answers, while the process rewards efficient tool use, parallelism, minimal unnecessary responses, and evidence-based claims. Evaluation uses Cursor Bench, built from real agent requests and curated optimal solutions, measuring correctness alongside adherence to existing abstractions and software engineering practices. The training stack combines asynchronous reinforcement learning with PyTorch and Ray, MXFP8 MoE kernels, expert parallelism, hybrid sharded data parallelism, and sandboxed cloud environments scaled to hundreds of thousands concurrently.

Takeaways

1. 1

   Cursor Bench evaluates agent usefulness using real requests and curated optimal solutions, including correctness, adherence to existing abstractions, and software engineering practices.

2. 2

   Reinforcement learning rewards efficient tool use and parallelism while encouraging minimal unnecessary responses and avoiding claims without evidence.

3. 3

   The training system supports thousands of NVIDIA GPUs and hundreds of thousands of concurrent sandboxed coding environments for tool-use training.