---
title: "A technical report on Composer 2"
description: "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."
---

# A technical report on Composer 2

[Cursor](https://yomu.fyi/company/cursor) · Sasha Rush · Mar 27, 2026

**Type:** Explainer

## 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.

## Context

Public coding benchmarks often use over-specified tasks, narrow solutions, and small codebases that do not reflect developers’ real work. The team built CursorBench from real coding sessions to evaluate terse or ambiguous prompts requiring changes across many files.

## Approach / What changed

Composer 2 uses continued pretraining on a code-emphasizing data mix followed by large-scale reinforcement learning in realistic Cursor sessions with the deployed model’s tools and harness. Its infrastructure includes custom low-precision kernels for MoE training on Blackwell GPUs, an asynchronous multi-region RL pipeline, and Anyrun for sandboxed coding environments.

## Takeaways

- Reducing pretraining loss improved downstream RL performance, with stronger base knowledge reliably translating into a better agent.
- CursorBench tasks come from engineering-team coding sessions and include terse, ambiguous prompts requiring hundreds of lines of changes across many files.
- 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.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [GPU](https://yomu.fyi/topic/gpu), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Reinforcement Learning](https://yomu.fyi/topic/reinforcement-learning)

- Source: [Cursor](https://cursor.com/blog/composer-2-technical-report)
- Source URL: https://cursor.com/blog/composer-2-technical-report
- Ingested by Yomu: 2026-09-02T11:42:46.232Z

[Read original post](https://cursor.com/blog/composer-2-technical-report)