Moonshot’s Kimi K3 beats Anthropic’s best on key benchmarks
The ninth World Artificial Intelligence Conference in Shanghai, which drew tens of thousands of visitors to a hall covering roughly 20 football fields of exhibition space, has become China’s premier event to showcase artificial-intelligence achievements. This year’s edition featured companies racing to close the gap with American rivals while pushing to deploy AI broadly across the economy, guided by a government blueprint called AI Plus that calls for embedding the technology in 90% of the economy by 2030, according to the Wall Street Journal.
The Journal reported that Moonshot AI’s new Kimi K3 model upended existing assumptions about the relative capabilities of Chinese and U.S. AI systems because it earned higher scores on many benchmarks than Anthropic’s Claude Opus 4.8, the most powerful model it likely could have accessed for distillation. Research has suggested a smaller model rarely surpasses the bigger model it distills, the Journal noted. In light of Kimi K3, some researchers including OpenAI executive Dean Ball have said that distillation—the practice of training a smaller model by extracting knowledge from a larger one—no longer explains the rapid advances of Chinese models.
Alibaba claimed to have surpassed Moonshot’s model with its own new AI release two days after Moonshot’s unveiling. An Alibaba business development manager told the Journal: “It’s fun to be at the frontier of the whole thing.”
Chinese AI models are gaining global traction, partly because many are open-source and generally cheaper to run than Western models, the Journal reported. The pattern mirrors industries such as smartphones and electric vehicles, which emerged in the West before China seized dominant market share.
“If it’s already happened in three industries, could that happen in AI as well? My answer would be, just given what’s already happening in AI today, it’s likely,” Mehran Gul, an author and speaker on technology, told the Journal.
The push for real-world applications was visible across the exhibition floor. Wang Xiaogang, chairman of Chinese firm Ace Robotics, said in an interview that building a sophisticated robot requires a long supply chain and many partners. “In China, they are very good at this,” he said.
At the booth for Suzhou-based UniX AI, a $44,000 humanoid robot completed household tasks such as making the bed, putting laundry in a washing machine and doing dishes. A marketing representative said the robot was already being used in hotels and restaurants, according to the Journal.
“Even though it might not be 100% mature right now, once it truly matures and stabilizes, it’s quite possible it could perform even better than humans,” he said.
The demonstrations were not always flawless. In one demo, the robot moved shakily and slowly, with laundry getting stuck on its claw-like hands. It once missed the dishwasher entirely, sending a plate clattering to the floor, the Journal reported.
Moonshot AI suspended new subscriptions for its premium services after demand for its Kimi K3 model strained its computing resources, a sign that U.S. sanctions on advanced chip technology continue to constrain some Chinese firms.
Alibaba executive Steven Hoi said the company’s main focus at the conference was to find “real-world agentic use cases where AI models can be used by many people.”
Attendees used paper fans and large blocks of ice placed by organizers to cool down from heat generated by computer servers, the Journal reported.