Composer 2.5

### TL;DR

Composer 2.5 is Cursor's latest AI coding model — a substantial improvement over Composer 2 in intelligence and behavior. It excels at long-horizon agentic tasks, follows complex instructions more reliably, and introduces targeted RL with textual feedback and 25x more synthetic training tasks.

Key Insights & Metrics

Key Features

• Targeted RL with textual feedback — localized training signal at specific trajectory points so the model learns which exact decision helped or hurt, not just the final reward
• 25x more synthetic tasks than Composer 2 — including feature deletion, byte code reverse engineering, and complex multi-file refactors for continual intelligence scaling
• Sharded Muon optimizer with dual mesh HSDP — asynchronous distributed orthogonalization enabling 0.2s optimizer steps on the 1T parameter model

→ Related Releases

Discussion