Cursor releases Composer 2.5 coding model with 25x larger synthetic training data
What happened
Cursor 把代码助手 Composer 升级到了 2.5 版，底层还是基于月之暗面的 Kimi K2.5 开源模型。这次主要做了三件事：一是用“文字反馈强化学习”，在模型犯错的地方直接插一句提示（比如“提醒：可用工具有这些”），让模型在那个点上学会纠正，而不是靠最后的总分去猜哪里做错了；二是把合成训练数据的量提到了上一代的 25 倍，并且动态生成更...
From AI HOT 精选
Coverage
Follow the reports to see the story from different sides.
- AI HOT (Curated Pool)PickCursor releases Composer 2.5, calling it its strongest model yet
Cursor released Composer 2.5, claiming a 10x efficiency gain at comparable capability, with larger training scale, more complex reinforcement-learning environments, and a text-feedback mechanism.
- AI HOT (Curated Pool)PickCursor releases Composer 2.5 coding model
Cursor released Composer 2.5, claiming up to 10x higher efficiency on long coding tasks; the model is further trained on Moonshot’s Kimi K2.5 and uses text feedback for 100k-token-scale trajectories.
- AI HOT (Curated Pool)PickComposer 2.5 release and technical analysis
Cursor released Composer 2.5, built on a Moonshot open-source checkpoint, trained with synthetic data from real codebases at 25 times the previous scale, and updated with text-feedback reinforcement learning and a sharded Muon optimizer.