China's 4-Year-Old AI Startup Moonshot Outperforms Korea's National AI Models

Technology|
|
By Kim Sung-tae
|
Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends] - Seoul Economic Daily Technology News from South Korea
Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends]

As the era of AI agents—artificial intelligence that autonomously makes decisions and performs tasks—gains momentum, Chinese AI startups including Moonshot are posting benchmark scores that surpass global Big Tech companies. Analysts attribute this to a combination of key talent returning from the United States and China's robust government support policies. Meanwhile, models participating in Korea's government-led "Independent AI Foundation Model Project" still show significant gaps with global leaders, prompting calls for aggressive talent recruitment policies and improved research and development environments.

According to IT industry sources on Wednesday, Moonshot's AI model "Kimi K2.5," released on January 27, scored 50.2 on the HLE benchmark, which evaluates the ability to solve complex reasoning problems using tools. This surpasses competing models including Gemini 3 Pro (45.8) and GPT-5.2 (45.5). HLE is considered one of the key indicators for measuring AI agent capabilities.

Kimi K2.5 is particularly noted for its high utility as an AI agent, reflected in its strong benchmark scores. Moonshot emphasized that Kimi K2.5 utilizes "agent swarm" technology, where a leader AI agent breaks down tasks and distributes them to up to 100 subordinate AI agents that execute the work in parallel—similar to a team leader delegating tasks to team members.

Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends] - Seoul Economic Daily Technology News from South Korea
Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends]

Kimi K2.5's overall performance is also rated highly. The model scored 47 points on Artificial Analysis's Intelligence Index, ranking fifth overall and matching performance levels of major U.S. AI companies including Google, OpenAI, Anthropic, and xAI.

Moonshot was founded in 2023 by Yang Zhilin, born in 1992, who previously worked at Big Tech research organizations including Google Brain and Facebook AI Research. According to CNBC and other foreign media, the company's valuation has reached approximately $4.8 billion (about 6.93 trillion won). The startup has attracted investment from China's Alibaba and Tencent. Yang, originally from Guangdong Province in southern China, graduated top of his class from Tsinghua University's computer science department before earning his doctorate in computer science from Carnegie Mellon University in the United States. After working at Facebook and Google, he returned to China and currently serves as an assistant professor at Tsinghua University.

Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends] - Seoul Economic Daily Technology News from South Korea
Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends]

Alibaba Group's "Qwen3 Max Thinking" is also drawing attention. Qwen3-Max-Thinking is a reasoning AI equipped with more than 1 trillion parameters that has demonstrated strong performance in agent-type AI capabilities, scoring 49.8 on HLE and outpacing competing models.

Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends] - Seoul Economic Daily Technology News from South Korea
Korea's 'National AI' Overtaken by 4-Year-Old Chinese 'Startup' Moonshot... What's the Breakthrough? [Kim Sung-tae's Deep Tech Trends]

The Qwen3-Max-Thinking model is designed to autonomously select search, memory, and code execution functions based on conversation flow to solve problems. It also applies an "advanced test-time scaling" approach that flexibly increases the time and computational resources AI uses for reasoning, combined with a "multi-round test-time scaling strategy" that efficiently utilizes previous reasoning results to reduce repetitive inference and improve accuracy.

In contrast, Korean companies that advanced to the second round of the government's flagship "Independent AI Foundation Model Project" recorded relatively low scores. On the HLE Text-only evaluation, which assesses the ability to solve complex problems using only the AI model's own knowledge and reasoning without tools, LG AI Research's "K-EXAONE" scored 13.6, Upstage's "Solar Open" scored 10.5, and SK Telecom's (017670.KS) "A.X K1" scored 7.6. By comparison, Kimi K2.5 scored 31.5, while the top-tier Gemini 3 Pro Preview scored 37.7.

To catch up, voices are emerging that the government's AI promotion policies should focus on nurturing the overall domestic AI ecosystem. In particular, analysts argue that aggressive talent recruitment policies are needed, along with continuous improvement of R&D environments that relatively lack resources such as GPUs and data, and support to compensate for lower compensation levels compared to Big Tech. Additionally, there are calls to improve visa systems so that outstanding foreign talent can conduct stable research activities in Korea.

Kim Young-mu, an investor at Kakao (035720.KS) Ventures, said, "Looking at top-tier AI conferences in the United States, Korean researchers make up the largest share after American and Chinese researchers, showing that Korea has a deep talent pool." He added, "However, relatively few personnel want to continue their research activities based in Korea."

Kim noted, "In contrast, Chinese researchers have a strong desire to return to their home country after gaining experience in the United States. This is not simply a matter of patriotism but because they are supported by solid research infrastructure and the government's exceptional support policies that make it advantageous to continue research in China." He emphasized, "Korea must also provide outstanding talent with 'an environment where they can focus solely on research' and 'clear incentives' to bring about the practical return of Korean researchers."

The Software Policy Research Institute stated, "We need to link talent capabilities to domestic growth engines through return programs and the establishment of a global Korean AI talent network," adding that "when top-tier talent returns, the public and private sectors should cooperate to provide concentrated support including research funding, startup capital, and settlement assistance." The institute also said, "For talent residing overseas, we should encourage contributions to Korea without physical relocation through joint research or advisory activities."

Meanwhile, the government plans to recruit a total of 2,000 outstanding overseas researchers and emerging researchers to Korea by 2030, with 70% to be filled by overseas Korean scientists returning home. Additionally, the Ministry of Science and ICT plans to establish "AI-specialized pilot cities" with significantly reduced regulatory intensity for AI development and service demonstration to attract global talent and outstanding startups. The ministry is consulting with the Ministry of Foreign Affairs and Ministry of Justice on measures including operating a visa-free system for a certain period so that AI talent can easily enter the country and engage in activities without worrying about residency status.

Related Video

AI-translated from Korean. Quotes from foreign sources are based on Korean-language reports and may not reflect exact original wording.