# Cursor Launches Self-Developed Agentic Coding Model Composer 2.5

- Category: Flash (快讯)
- Published: 2026-05-19T11:58:00.000Z
- Source date: 2026-05-19
- HTML: https://www.88z.ai/en/news/flash-20260519-47e76ffe67967175
- Markdown: https://www.88z.ai/en/news/flash-20260519-47e76ffe67967175.md
- Other language: https://www.88z.ai/news/flash-20260519-47e76ffe67967175.md

Elon Musk invited users on X to test Cursor's latest model, Composer 2.5, which was partially trained using Colossus 2. The official statement indicates it builds on Moonshot AI's Kimi K2.5, focusing on improving stability for long tasks and adherence to complex instructions, while enhancing performance through directed reinforcement learning and expanded synthetic training.

## At a glance

- Composer 2.5 is trained on Kimi K2.5 with partial use of Colossus 2, focusing on long-task stability and complex instruction adherence.
- It employs text-feedback directed reinforcement learning, expands synthetic training scale to 25 times the predecessor, and uses sharded Muon and dual-grid HSDP for optimization.

## The update

Elon Musk recently issued an invitation on the X platform, urging users to participate in testing Cursor's latest programming model, Composer 2.5. It is reported that the training process of this model partially utilized Colossus 2 computing resources. Cursor officially disclosed that Composer 2.5 is trained based on Moonshot AI's Kimi K2.5, with its core optimization focus on improving stability during long-cycle tasks and enhancing adherence to complex instructions. In terms of technical implementation, the new model adopts a directed reinforcement learning method based on text feedback. Furthermore, the scale of its synthetic training data has been significantly expanded, reaching twenty-five times that of the previous generation. To ensure training efficiency, the team also introduced sharded Muon algorithms and dual-grid HSDP technology for optimization, thereby supporting larger-scale model training and iteration.

## Sources

- [IT之家](https://www.ithome.com/0/952/106.htm) (2026-05-19)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.