AI Agents & Agentic Engineering
  • Context Engineering
  • Building AI Agents
  • Agentic Engineering Best Practices
  • Auto-Research and Self-Improving Agents
  • Self-Maintaining Agents
  • Scaling Long-Running Autonomous Coding
  • Data Agents
  • Agent Memory and Harness Ownership
LLM Internals & Training
  • Reinforcement Learning for LLMs
  • Cursor Composer 2
  • Training-Free GRPO
  • Prompt Repetition for Non-Reasoning LLMs
  • LLM Quantization
  • LLM Knowledge Bases
AI Safety & Security
  • Constitutional Classifiers
  • Project Glasswing
Graphics & Low-Level Programming
  • Shaders and GPU Programming
Personal Development & Mindset
  • Deep Focus and Productivity
  • Identity and Life Design
  • Self-Competition and Entrepreneurship

Cursor Composer 2

Cursor's Composer 2 is a purpose-built coding model created through continued pretraining and reinforcement learning, designed to closely emulate the Cursor IDE environment.

Source: Cursor AI thread, Composer 2 Technical Report
Attachment: Composer 2 Technical Report.pdf

Three Main Efforts

Composer 2 overview

1. Continued Pretraining

Continued pretraining improvements

Starting from a base model (Kimi K2.5), continued pretraining on coding-specific data produced consistent improvements in downstream coding performance. The pretraining phase focuses on building strong coding fundamentals before the RL phase refines behavior.

2. Reinforcement Learning

RL phase results

The RL phase is described as "critical for final performance." Key findings:

  • Simple approaches often work best and improve performance broadly
  • Training was distributed across 3-4 clusters around the world using Fireworks' multi-region RL infrastructure
  • Environment scaling for RL was a significant engineering challenge

3. Benchmark Development (CursorBench)

CursorBench results
Cursor developed an internal benchmark called CursorBench representing a more realistic sampling of coding problems. The motivation: existing benchmarks don't include the complex problems software engineers encounter daily. CursorBench aims to bridge this gap.

Infrastructure

The technical report covers:

  • Custom kernels developed and open-sourced for the project
  • Distributed training infrastructure
  • Environment scaling for RL
  • Collaboration with Fireworks and Colfax

Built on top of contributions from Kimi K2.5, Ray, ThunderKittens, PyTorch, and other open-source communities.

Significance

Composer 2 demonstrates that continued pretraining + RL can produce a coding model that is both highly capable and specifically tuned for the IDE environment. The RL phase, powered by distributed infrastructure (see reinforcement-learning-for-llms), was the critical differentiator for final performance.

This model powers the agentic coding capabilities used in scaling-autonomous-coding, where Cursor runs hundreds of concurrent agents for weeks.

Related Topics

  • reinforcement-learning-for-llms - The infrastructure that enabled distributed RL training
  • scaling-autonomous-coding - How Cursor uses these models for multi-agent work
  • llm-quantization - Related optimization techniques for model efficiency