Moonshot AI’s newest release, Kimi K2.5, has narrowed the gap between the U.S. and China to the closest it has ever been in the development of artificial intelligence (AI) models, according to third-party evaluations. That feat by the Beijing-based start-up, founded in March 2023, has raised questions about the efficacy of U.S. policies – most notably, export controls on advanced semiconductors – to constrain China’s AI development, said Kyle Chan, Fellow of the John L. Thornton China Center at the Washington-based think tank Brookings Institution. “It seems that, for these Chinese start-ups, it was just a matter of getting access to capital,” Chan said. “I wouldn’t have guessed that Chinese AI companies would continue to keep pace with their U.S. peers as recently as a month or two ago.” Moonshot was valued at USD4.3 billion after raising USD500 million in its latest Series C funding round in December, according to venture capital tracker ITJuzi. Investors include IDG Capital, Tencent Holdings and Alibaba Group Holding.
Kimi K2.5 scored just a handful of points behind the leading models of U.S. heavyweights OpenAI, Anthropic and Google DeepMind in comprehensive performance evaluations by Artificial Analysis, the third-party benchmarking firm said. Like other Chinese AI firms, Moonshot publicly released the source code of its new model, meaning that developers around the world with the requisite hardware could freely download and customize the powerful 595-gigabyte model without paying for expensive subscriptions. The open-source nature of such a powerful model challenges the strategy of American AI companies that develop closed-source models, according to Chan. “Especially when it comes to subscription-based start-ups like OpenAI and Anthropic, when you have a Chinese model that is basically just as good and open source, it seems to be a real challenge to their business models,” he said.
Raw intelligence aside, Kimi K2.5 was also much cheaper to run than its U.S. counterparts. Artificial Analysis estimated that Kimi K2.5 was more than four times cheaper than the leading models of OpenAI and Anthropic. According to Moonshot, Kimi K2.5 was built on the same Mixture-of-Experts architecture that has helped drive efficiency gains for other Chinese AI companies over the past year. U.S. restrictions on access to computational resources have forced these companies to maximize the cost efficiency of their models. Moonshot’s new release marked its first multimodal model, which means it can process image and video in addition to text. That removed a “critical barrier” to the adoption of open models, as other leading Chinese models have yet to add this feature, according to Artificial Analysis, as reported by the South China Morning Post.
