As a Chinese artificial intelligence company unveiled an ultra-large-scale AI model boasting GPT-6-level performance, concerns have emerged over the possibility of a "second DeepSeek shock." South Korea, which is striving to catch up with China's leading AI models, faces growing pressure to narrow the technological gap.
According to industry sources on Monday, Chinese AI firm Moonshot AI recently released its ultra-large-scale AI model, "Kimi K3.1 Pro." The model features a total of 2.8 trillion parameters and can process up to 1 million tokens at once. By employing a Mixture of Experts (MoE) approach, it activates only approximately 3.7% of the total parameters—104 billion—for actual computation.
In benchmark results from Artificial Analysis, a global AI evaluation platform, Kimi K3.1 Pro scored 55.1 points on the Intelligence Index. This represents an increase of 11.1 points over its predecessor, Kimi K3 (44 points), placing it fourth overall. The score is higher than that of OpenAI's GPT-6 Astra (53 points).
In contrast, major domestic AI models have seen significant drops in scores following recent changes to Artificial Analysis's evaluation criteria. Motif Technologies' "Motif 3" scored 34 points, Upstage's "Solar Pro 4" recorded 27 points, SK Telecom's "A.X-K2" achieved 21 points, and LG AI Research's "K-Exaone 2.0" logged 20 points. Even Motif 3, which holds the highest score among domestic models, trails Kimi by 21.1 points.
Kimi also demonstrated competitiveness in coding performance. According to Moonshot AI's internal evaluation, it scored 65.7 points on "TerminalBench 4.0," surpassing GPT-6 Astra (58.2 points). The model also offers price competitiveness. Kimi's API (Application Programming Interface) usage fee is $2 for input and $10 for output per 1 million tokens, which is half the cost of Claude Opus 5.5. Moonshot AI plans to release the model weights by the 4th of next month. Once the weights are made public, other companies will be able to run Kimi on their own servers or utilize it through additional training.
The rapid advancement of Chinese AI is also weighing on South Korea's catch-up strategy. The government is pushing forward a Frontier AI project worth 470 billion won from 2027 to 2031. While the goal is to catch up with China's leading open-weight (model weight disclosure) AI, Chinese models have already risen to a level competing with top-tier U.S. systems before the project has fully ramped up.
Professor Mo Jeong-hoon of the Department of Industrial Engineering at Yonsei University stated, "China is already closely pursuing the United States, but the gap for South Korea is significant." He added, "Even if it is difficult to catch up immediately, an independent AI model can serve as a means to enhance negotiating power, much like how Samsung Electronics secures bargaining leverage with Qualcomm through its own application processor (AP), Exynos."
Lee Seong-yeop, a professor in the Technology Management program at Korea University, advised, "Resources such as GPUs and talent must be concentrated, and decision-making processes should be streamlined." He suggested, "Rather than pursuing multiple AI development projects separately, strategies should be considered that integrate efforts around Frontier AI."
A government official revealed, "There is a sense of crisis that we must move quickly because AI companies are evolving rapidly at intervals of nearly three months." The official added, "Large-scale investment is needed to catch up with Chinese open-weight models."
in TECH
