Composer 2
Properties
created
04.04.2026, 12:00
modified
26.07.2026, 13:47
published
Empty
sources
Composer 2 Technical Report
topics
Agentic Software Engineering, Model Training, Benchmarks
authors
Empty
ai-assisted
Yes
- Specialized model by Cursor Research for agentic software engineering
- Focus: long-term planning and coding intelligence in multi-step, realistic dev environments
- 54 authors; March 25, 2026
- Paper: Composer 2 Technical Report
# Training
- Two-phase approach:
- Continued pretraining on code and software engineering data
- Large-scale reinforcement learning on realistic coding problems
- Training infrastructure mirrors the actual Cursor deployment harness — same tools, environments, and structure used in production
# Benchmarks
| Benchmark | Score |
|---|---|
| CursorBench | 61.3 |
| Terminal-Bench | 61.7 |
| SWE-bench Multilingual | 73.7 |
- CursorBench — internal benchmark derived from real software engineering challenges faced by Cursor users; measures practical agentic coding ability
- Terminal-Bench — evaluates terminal-based software engineering tasks
- SWE-bench Multilingual — real GitHub issues across multiple programming languages
# Key Ideas
- Training on environments that closely match real problems (not toy tasks) is central to the model’s effectiveness
- RL with execution feedback allows the model to learn from actual code running — not just human labels
- Significant improvement over prior Composer versions on all benchmarks