Artificial Intelligence

Cursor launches Composer 2.5 model for long-running AI coding tasks at cheaper cost

North America / United States0 views1 min
Cursor launches Composer 2.5 model for long-running AI coding tasks at cheaper cost

Cursor, a US-based AI coding startup, launched Composer 2.5, a model trained for long-running coding tasks at a lower cost, built on the open-source Kimi K2.5 base from Chinese startup Moonshot AI. The company also announced plans to collaborate with SpaceX’s AI division to train a significantly larger model using its Colossus 2 supercomputer, amid a reported $60 billion acquisition deal by SpaceX." "article": "Cursor, an AI coding startup, unveiled Composer 2.5 on May 18, 2026, designed for sustained coding tasks with improved instruction-following and cost efficiency. The model builds on Kimi K2.5, an open-source base from Moonshot AI, a Chinese firm backed by Alibaba and HongShan. Cursor acknowledged earlier criticism over its Composer 2 model’s origins, stating only 25% of its compute came from the base, while the rest was proprietary training. Composer 2.5 introduces targeted textual feedback during reinforcement learning to refine behavior, such as correcting tool-use errors with hints like ‘Reminder: Available tools…’. The model was trained on 25 times more synthetic coding tasks than its predecessor but risks reward hacking due to synthetic data reliance. Cursor mitigates this via agentic monitoring tools. The company also doubled Composer 2.5’s free usage for a week and revealed plans to develop a ‘significantly larger model’ with SpaceX’s AI division, SpaceXAI. This follows SpaceX’s April announcement of a potential $60 billion acquisition of Cursor, pending regulatory approval. The collaboration will leverage SpaceX’s Colossus 2 supercomputer, featuring millions of H100-equivalent GPUs. Cursor’s growth highlights its $2.3 billion funding round last year, valuing the firm at $29.3 billion, with annual revenue reportedly exceeding $2 billion. The partnership with SpaceXAI underscores the intensifying AI competition, particularly between US and Chinese tech ecosystems. Composer 2.5’s release marks a step toward Cursor’s long-term goal of building fully proprietary models while addressing ethical and technical challenges in AI development.

Cursor, an AI coding startup, unveiled Composer 2.5 on May 18, 2026, designed for sustained coding tasks with improved instruction-following and cost efficiency. The model builds on Kimi K2.5, an open-source base from Moonshot AI, a Chinese firm backed by Alibaba and HongShan. Cursor acknowledged earlier criticism over its Composer 2 model’s origins, stating only 25% of its compute came from the base, while the rest was proprietary training. Composer 2.5 introduces targeted textual feedback during reinforcement learning to refine behavior, such as correcting tool-use errors with hints like ‘Reminder: Available tools…’. The model was trained on 25 times more synthetic coding tasks than its predecessor but risks reward hacking due to synthetic data reliance. Cursor mitigates this via agentic monitoring tools. The company also doubled Composer 2.5’s free usage for a week and revealed plans to develop a ‘significantly larger model’ with SpaceX’s AI division, SpaceXAI. This follows SpaceX’s April announcement of a potential $60 billion acquisition of Cursor, pending regulatory approval. The collaboration will leverage SpaceX’s Colossus 2 supercomputer, featuring millions of H100-equivalent GPUs. Cursor’s growth highlights its $2.3 billion funding round last year, valuing the firm at $29.3 billion, with annual revenue reportedly exceeding $2 billion. The partnership with SpaceXAI underscores the intensifying AI competition, particularly between US and Chinese tech ecosystems. Composer 2.5’s release marks a step toward Cursor’s long-term goal of building fully proprietary models while addressing ethical and technical challenges in AI development.

This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.

Comments (0)

Log in to comment.

Loading...