SOURCE / COMPANIES
Chinese AI start-up unveils latest reasoning model comparable to OpenAI o1 series
Published: Nov 17, 2024 10:04 PM
AI Photo: VCG

AI Photo: VCG


Artificial general intelligence start-up Kimi, owned by Chinese AI start-up Moonshot AI, on Saturday launched its first reasoning AI model k0-math. The model can be compared to US-based OpenAI's reasoning AI series o1, including the o1-mini and o1-preview, in certain mathematics tests.

In four mathematical benchmark tests - China's high school entrance examination, college entrance examination, postgraduate entrance examination and math with introductory competition problems - the k0-math initial model outperformed OpenAI's o1-mini and o1-preview models, according to a statement sent from Moonshot AI to the Global Times on Sunday. 

In the two more challenging competition-level math problem sets - OMNI-MATH and AIME tests - the performance of the k0-math initial model reached 90 percent and 83 percent of the highest scores achieved by o1-mini, respectively.

The k0-math is equipped with enhanced learning and chain-of-thought reasoning technologies, bringing more powerful reasoning abilities. It significantly improves the capability to solve complex mathematical problems by simulating the thinking and reflection processes of the human brains, said the statement. 

However, k0-math still has limits in solving some geometry problems that are difficult to be described by using the LaTex format, and leads to "overthinking" of simple mathematics problems such as 1+1=?, as the developer pointed out that it requires better generalization to be applied in more scenarios across various disciplines.

Chinese developers have been popularizing various types of AI tools in different utilization scenarios. November 16 marked the first anniversary of AI chatbot Kimi's release to public, and its monthly active user volume across multiple platforms surpassed 36 million in October, said the statement. 

In August, the Chinese text-to-video AI model Vidu was made available to users, featuring core functionalities of generating videos and images from texts. It offers users the choice of 4- or 8-second video clips, with a resolution of up to 1080P. 

Global Times