Sasha Rush, an ML researcher at Cursor, outlined training Composer 2 from the Kimi K2.5 base model with 1 trillion total parameters through continued pre-training and long-horizon reinforcement learning

— The base model was chosen for infrastructure compatibility and benchmark performance.

Original post

OPSR#23Sasha Rush|@SRUSH_NLP

Talk: Training Composer https://www.youtube.com/watch?v=uTgqYeVxy2c Overview of the methods that we use at Cursor to build our model.

12:05 PM · May 21, 2026 View on X

QUOTE POSTWZ#584Wenting Zhao|@WZHAO_NLP

Listening to Sasha explaining things is one of my favorite things 😍

SRSasha Rush@srush_nlp

Talk: Training Composer https://www.youtube.com/watch?v=uTgqYeVxy2c Overview of the methods that we use at Cursor to build our model.

7:05 PM · May 21, 2026 · 34.1K Views

3:51 AM · May 22, 2026 · 1.8K Views