In this workshop, Sasha Rush walks through how Cursor's research team builds Composer 2 - from base model choice to long-horizon reinforcement learning.
Key topics covered:
Base model selection:
Composer 2 starts from Kimi K2.5 (1T params, 32B active, 256K context). The choice came down to both infrastructure fit and initial benchmark scores.
Continued pre-training:
A coding-focused pre-training stage builds domain knowledge. More tokens here translate into measurably higher rewards after the RL stage.
Long-horizon RL with auto-install:
Composer 1.5 bootstraps each training environment by exploring the repo, generating install commands, and writing verification tests before RL begins.
Reward shaping and self-summarization:
A nonlinear length penalty balances speed and depth, while self-summarization lets the model continue past its context limit and still share one final reward across the rollout.
Cursor Bench:
An internal eval of short, ambiguous prompts and large multi-file diffs from real engineer queries. It separates strong from weaker models much more cleanly than SWE-bench.
-----------------------
Helpful resources:
Models: https://cursor.com/docs/models
Cursor docs: https://cursor.com/docs
-----------------------
This is a recording from our live session on May 14, 2026.
For more events like this, check out upcoming workshops at https://cursor.com/workshops
