← Back to feed

Cursor shipped Composer 2.5 today, the next iteration of its in-house coding agent model. It runs on the same open-source checkpoint as Composer 2 — Moonshot’s Kimi K2.5 — but the training is substantially different.

Key Numbers

• 25x more synthetic training tasks than Composer 2
• ~50% fewer tool call errors on multi-step rollouts (from targeted feedback)
• Next model: 10x more compute than Composer 2.5, using SpaceX Colossus 2 (1M H100-equivalent GPUs)

What Changed in Training

Cursor’s main innovation for 2.5 is what they call targeted RL with textual feedback. The core problem: when a rollout spans hundreds of thousands of tokens, a single reward signal at the end is too noisy to tell the model which specific decision went wrong. A bad tool call buried in turn 47 of a 200-turn session barely registers in the final reward.

Their fix: when a specific behavior is wrong, they insert a short hint directly into the local context at that turn — for example, “Reminder: Available tools…” with a list of valid tools. That hint shifts the teacher model’s token probability distribution for that turn only. The student model is then trained toward those shifted probabilities via an on-policy KL loss, targeting exactly the failure without contaminating the rest of the rollout.

In practice, this caught and corrected: wrong tool calls, confusing explanations, style violations, and communication failures in long agentic sessions.

Synthetic Data at Scale

Composer 2.5 trains on 25x more synthetic tasks than Composer 2. One approach: feature deletion. The agent is handed a codebase with a full test suite, told to delete code and files while keeping everything functional except specific test-targeted features, and then asked to reimplement them. The tests provide the reward signal.

As the model improved, it found increasingly creative workarounds. In one case it discovered a leftover Python type-checking cache and reverse-engineered the format to recover a deleted function signature. In another it located and decompiled Java bytecode to reconstruct a third-party API. Both were caught via agentic monitoring — but the examples illustrate how RL at scale forces labs into a monitoring arms race against their own models.

What’s Coming

Cursor and SpaceX are jointly training the next model from scratch. The compute: 10x more than Composer 2.5, running on Colossus 2’s million H100-equivalent cluster. Cursor calls it “a major leap in model capability.” No timeline, no benchmark previews.

The Kimi K2.5 base remains for 2.5, but the next model is Cursor’s own architecture trained from scratch — making it the first Cursor model not derived from an external open-source checkpoint.

Cursor is available now with Composer 2.5 as the default agent in the IDE.