Kimi Releases Math Model k0-math, Math Performance Beats o1; Yang Zhilin Discusses Retention and Scaling Law

On November 16, on the first anniversary of Kimi's full launch, Moonshot AI launched k0-math, a reasoning-enhanced math model, along with a series of new features for Kimi Explorer Edition. The company says the newly launched k0-math math model outperforms o1 on tests.

According to the official introduction, on four math benchmarks—including the middle school entrance exam, the college entrance exam, the graduate entrance exam, and MATH for introductory competition problems—the first-generation k0-math model outperformed o1-mini and o1-preview. In October 2024, the Kimi intelligent assistant had more than 36 million active users across all platforms, and new features of k0-math and Kimi Explorer Edition will be rolled out to the web and app.

As for why it chose math scenarios, Kimi founder Yang Zhilin believes it is the most suitable scenario for training AI's thinking ability, a process of continuous thinking and trial and error that does not require interaction with the outside world. He also pointed out that the current version of k0-math cannot yet solve geometry-type problems that are difficult to describe in LaTeX format, will overthink simple problems, and in some scenarios will guess answers.

Regarding the industry discussion about the Scaling Law hitting a wall, Yang Zhilin said: 'All good algorithms are friends with Scaling. If your algorithm can unlock the potential of Scaling, it will continue to get better.' He also mentioned that the most important capabilities for AI next are thinking and interaction; thinking is far more important than interaction, thinking determines the upper limit, and interaction is a necessary condition.

On product and growth, Yang Zhilin said: 'For us, the most core thing is to do a good job in retention and Growth. Appropriate paid promotion is needed.' ^[3]