A new reasoning model developed by a Chinese artificial intelligence start-up – with its performance exceeding OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 in a number of metrics – has fanned fresh debate about another DeepSeek moment and the trajectory of America’s AI supremacy.
Beijing-based Moonshot AI, a start-up valued at US$3.3bil (RM13.64bil) and backed by Chinese tech giants like Alibaba Group Holding and Tencent Holdings, has presented another David-vs-Goliath story after creating an open-source model that “set new records across benchmarks that assess reasoning, coding and agent capabilities”.
The new reasoning model, Kimi K2 Thinking, was the most popular model for developers on Hugging Face as of Monday, while its release post on X had attracted 4.5 million views. The popularity of the model – a variant of the Kimi K2 model – had further grown after CNBC reported its training cost was merely US$4.6mil (RM19.01mil). Moonshot AI did not comment on the cost.
Even without factoring in its costs, the latest model has impressed the AI community. Thomas Wolf, co-founder of Hugging Face, commented on X that Kimi K2 Thinking was another case of an open-source model passing a closed-source model.
“Is this another DeepSeek moment?” Wolf asked. He was referring to the launch of the low-cost but high-efficient R1 reasoning model by the Hangzhou-based Chinese AI start-up earlier this year that had shaken the perception of absolute American AI supremacy. “Should we expect [one] every couple of months now?”
Kimi K2 Thinking scored 44.9% on Humanity’s Last Exam – a large language model (LLM) benchmark consisting of 2,500 questions across a broad range of subjects – exceeding GPT-5’s 41.7%, according to the company’s GitHub blog post.
It also outperformed US models in two specific benchmarks, including the BrowseComp benchmark, which evaluates the web browsing proficiency and information-seeking persistence of LLM agents, and the Seal-0 benchmark, designed to challenge search-augmented LLMs on real-world research queries.
The cost of Kimi K2 Thinking’s application programming interface was six to 10 times cheaper than those of OpenAI and Anthropic’s models, according to the Post’s calculation.
Zhang Ruiwang, a Beijing-based information technology system architect working in the internet sector, said the trend was for Chinese companies to keep costs down. “The overall performance of Chinese models still lags behind top US models, so they have to compete in the realms of cost-effectiveness to have a way out,” Zhang added.
Zhang Yi, chief analyst at consultancy iiMedia, said the training costs of Chinese AI models were seeing a “cliff-like drop” driven by innovation in model architecture and training technique, and input of quality training data. It marked a shift away from the heaping of computing resources in the early days.
“Striving to continue lowering [training and use] costs has become a key tactic for Chinese AI developers to attract developers in a heated competition,” Zhang said. – South China Morning Post
