Composer 2.5 Made the Cursor Data Advantage Visible
Insight
Composer 2.5 sits Pareto-dominant on the cost/quality frontier after only three to four weeks of RL on Colossus 2, using the same base model (Kimi K2.5) as Composer 2. The jump isn't a bigger model or longer pre-training — it's that Cursor allegedly has more tokens of coding data than the entire public internet, and RL on those tokens is enough to move the frontier on its own. This is the cleanest existence proof yet that proprietary coding tokens, applied via RL, beat scale alone in a specific domain.
Source Context
"Cursor's composer 2.5 model came out this week. And I mean, this is Pareto dominant. And this is just you know, three, four weeks of doing reinforcement learning on Colossus 2 with Cursor's data... Cursor allegedly has more tokens of coding data than exist on the public internet." — 49:39(https://www.youtube.com/watch?v=HGbA6ze0_3M&t=2979s)
"Composer 2.5 is the same base model as composer 2, which is Kimmy K K 2.5. Like, this is amazing. This is three or four weeks, and it is Pareto dominant." — 50:40(https://www.youtube.com/watch?v=HGbA6ze0_3M&t=3040s)