Kimi. VCG

Downloads of the new artificial intelligence (AI) model Kimi K2 doubled in a few days after Chinese tech startup Moonshot AI released an update to its Kimi reasoning model on July 11, drawing growing attention from multiple media outlets and global AI community, with some of them saying that it is "another DeepSeek moment."

On AI and machine-learning developer platform Hugging Face, downloads of Kimi K2 soared to 145,000 on Monday from 76,000 on Friday, the South China Morning Post (SCMP) reported on Tuesday.

The figures came after Beijing-based AI startup released its updated AI model on July 11, which became the world first open source trillion parameter Mixture of Experts (MoE) model.

MOE is a machine-learning approach that divides an AI model into separate sub-networks, or experts - each focused on a subset of the input data - to jointly perform a task. This is said to greatly reduce computation costs during training and achieve faster performance during inference.

"Moonshot's Kimi K2 soars in popularity amid experts' praise for Chinese AI developments," according to the SCMP report.

The release underscores China's accelerating open-source momentum. Nathan Lambert, a researcher at the Allen Institute for AI, posted on the social-media site Bluesky on July 14 saying that K2 is "the new best-available open model by a clear margin," while the model has already vaulted to the top of the LMSYS Chatbot Arena leaderboard.

In a Nature article published on July 16, the scientific journal hailed K2 as "another DeepSeek moment," referencing the shockwaves sent by DeepSeek's R1 model in January. The performance of Kimi K2 matches or surpasses that of Western rivals, as well as some DeepSeek models, across various benchmarks. Researchers around the world are now increasingly enthusiastic about the emergence of China's second high-performance AI model, it said.

The article highlights that Kimi K2 particularly excels at coding, achieving high scores in tests such as LiveCodeBench, a method for evaluating AIs that challenges models on code-related tasks.

Moonshot AI's Kimi chatbot was the third-most-used in China by November, according to Counterpoint, a marketing-research firm in Hong Kong. 

According to its official WeChat account, Kimi K2 has a total parameter count at the trillion scale (1T). However, it employs an MoE architecture, activating only 32 billion parameters at a time. It can independently orchestrate multi-step workflows by invoking external tools — browsing the Web, running code, or calling mathematical software — capabilities previously locked behind proprietary paywalls.

Similar to DeepSeek models, Kimi K2 is open-weight, meaning it can be freely downloaded and modified by researchers. 

The Nature article also suggested that the launch of a second high-performance model within six months indicates that China's recent AI breakthroughs are part of a sustained trend rather than isolated successes.

The widespread attention on Moonshot's model update signaled that China's open-source surge is anything but a one-off fluke, Tian Feng, president of the Fast Think Institute and former dean of Chinese AI software giant SenseTime's Intelligence Industry Research Institute, told the Global Times on Tuesday, adding that it also reflected the growing global recognition and influence of Chinese AI innovation, though facing relentless US crackdown.

"The reports of the media outlets not only affirm the country's technical strength, but also demonstrate the rising global competitiveness of Chinese AI companies, which could reshape the global AI landscape," Tian added.

China's influence in the field of AI has risen significantly, according to a latest report from the China Internet Network Information Center. As of March, a total of 346 generative AI services had completed registration with the Cyberspace Administration of China, said the report.

Global Times