Search

Search IconIcon to open search

Composer 2

Last updatedUpdated: by Jakub Žovák · 1 min read

Properties
tags cscs/swe
created 04.04.2026, 12:00
modified 26.07.2026, 13:47
published Empty
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

BenchmarkScore
CursorBench61.3
Terminal-Bench61.7
SWE-bench Multilingual73.7
  • 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