Cursor drops Composer 2.5 with Kimi base

This release marks a departure from agentic wrappers toward deeply specialized models that outperform general-purpose frontier LLMs in coding.

• –Massive synthetic training scale (25x) specifically targets 'feature deletion' and multi-file refactoring scenarios.
• –New textual feedback loop in RL directly addresses credit assignment for complex, long-running engineering plans.
• –Transitioning to a Kimi K2.5 base with 85% custom compute focus signals Cursor's intent to own the entire model-to-editor vertical.
• –Concurrent platform updates like parallel execution and native PR reviews further centralize the IDE as a proactive collaborator.
• –Benchmark results of 79.8% on SWE-Bench Multilingual place it at the absolute top of the field for production-ready agents.