On March 22, Cursor quietly released a new model, Composer 2, with the tagline "frontier-level coding intelligence." Soon after, an X user named Fynn revealed the truth: the model is essentially Kimi K2.5 with an added layer of reinforcement learning.
Kimi K2.5 is the open-source model from Moonshot AI, a Chinese company backed by Alibaba and Sequoia China (now renamed Hongshan). It has 1 trillion parameters.
Admitted Only When Asked
When Cursor launched Composer 2, its press materials contained no mention of Moonshot AI or Kimi.
After the revelation, Lee Robinson, Cursor's Vice President of Developer Education, commented:
Yes, Composer 2 started with an open-source base model! But in the final model, only about a quarter of the compute comes from that base model; the remaining three-quarters are from our own training.
His point was that Cursor performed extensive training on top of Kimi K2.5, so it should not be considered a mere wrapper.
Robinson also emphasized that the use was authorized—the official Kimi account on X confirmed it was a licensed commercial partnership through Fireworks AI.
Cursor co-founder Aman Sanger later stated: Not mentioning Kimi in the initial launch blog post was a mistake.
Why the Omission?
The reasoning is not hard to understand.
There is a strong narrative in the AI industry today framing US-China AI competition as a national struggle. Using a Chinese AI model as a product foundation is not politically favorable in the current climate.
A large portion of Cursor's user base consists of enterprise developers in the United States. Using the model quietly is fine, but actively promoting that the underlying model is Chinese—the market risk team apparently deemed that risk not worth taking.
The issue here is not technical. Kimi K2.5 is open-source and the licensing is fine. The problem is information asymmetry—users thought they were using Cursor's own frontier model, when in reality the foundation was someone else's.
A Larger Problem
Cursor's case actually reveals a common industry phenomenon: many self-proclaimed proprietary AI products are built on fine-tuned or reinforcement-learned versions of open-source models, but their launches tend to emphasize their own work while downplaying or omitting the open-source source.
This is not necessarily a technical problem (Cursor's three-quarters self-computed training is not false), but it raises a question about transparency.
On the other hand, this incident also speaks to the quality of Kimi K2.5: a company valued at billions of dollars, claiming to be a top-tier US AI coding tool, chose a Chinese open-source model as its technical foundation—that is perhaps the most genuine endorsement.
Moonshot AI is probably quite pleased.
Sources: Cursor admits its new coding model was built on top of Moonshot AI's Kimi (TechCrunch); CocoLoop, China's Moonshot releases a new open source model Kimi K2.5 and a coding agent (TechCrunch)