China's Kimi K3 narrows AI gap with US to just months, putting Korea on alert

A wave of AI shocks from China is rattling the global artificial intelligence market.

Following DeepSeek, Chinese AI startup Moonshot's latest model, Kimi K3, has demonstrated performance on par with flagship models from Anthropic and OpenAI — and analysts say the technology gap between China and the United States has now narrowed to as little as two months.

South Korea, which has set its sights on becoming one of the world's top three AI powers, is now on high alert. Analysts say the country must move faster to keep pace in an intensifying global AI race.

AI performance research firm Artificial Analysis rated Kimi K3 at 57 points on its intelligence index Monday, placing it fourth among all existing AI models — on a level with Anthropic and OpenAI's flagship offerings and ahead of Google's Gemini and xAI's Grok.

In the detailed rankings, Anthropic's Claude Fable 5 led at 60 points, followed by OpenAI's GPT-5.6 Sol Max at 59 and GPT-5.6 Sol XHigh at 58. Kimi K3 came in fourth at 57, ahead of xAI's Grok 4.5 (54 points), Meta's MuseSpark 1.1 (51 points) and Google's Gemini 3.5 Flash (50 points).

Kimi K3 is the world's largest open-weight AI model, with about 2.8 trillion parameters. In AI development, a higher parameter count generally allows a model to learn more complex content.

Anthropic's top-tier Claude model, Opus 4.8, is estimated to have more than 1.5 trillion parameters. Kimi K3 is regarded as the world's first open-source model in the 3 trillion-parameter class.

Moonshot said Kimi K3 is optimized not for simple question-and-answer tasks but for handling large-scale software repositories, operating terminal tools, and combining coding with visual feedback. The company added that the model's image and video comprehension capabilities allow it to modify digital content such as interfaces and games.

The global ICT industry sees Kimi K3 as a serious threat to US dominance in AI, largely because its competitive price-to-performance ratio could allow it to capture global market share rapidly.

Kimi K3's API is priced at $3 per million input tokens and $15 per million output tokens — not as cheap as earlier Chinese AI models, but a calculated bet that pairs global-tier performance with pricing that undercuts US rivals. GPT-5.6 Sol is priced at '$5 for input and $30 for output,' while Claude Opus 4.8 costs '$5 and $25.'

Kimi K3's arrival has also prompted analysis that the technology gap between China and the United States has narrowed to just two months.

Ion Stoica, a professor at the University of California, Berkeley, told Bloomberg that Chinese models used to lag US top-tier models by six to nine months, but the gap is now around two to three months.

South Korea is also on alert as China closes the AI gap following DeepSeek. Despite setting a goal of becoming one of the world's top three AI powers, the country has yet to produce a homegrown AI model that has made a mark on the global stage.

A key inflection point is expected to come with the evaluation process for the "Dokja AI Foundation Model Project," known as Dokpamo, with results expected as early as the end of this year.

The government-backed Dokpamo project, aimed at building sovereign AI capabilities, is set for a second round of evaluations in early August. LG AI Research, SK Telecom and Upstage have completed development of their respective models, while Motif Technologies, which joined the project in February, is set to finish development by the end of this month.

Once all four teams complete their models, a second evaluation will be held in early August to eliminate one team. The government plans to conduct a third evaluation within the year to select the final two teams.

The government has also outlined plans to accelerate AI infrastructure expansion with the goal of lifting South Korea's AI competitiveness to second in the world.

Bae Kyung-hoon, deputy prime minister and minister of science and ICT, said at a presidential briefing held at Cheong Wa Dae's banquet hall on Thursday that he expects the results of the second AI model evaluation in August to open the door to challenging for second place, beyond the current third-place standing.

sjpark@heraldcorp.com