▲ China's Moonshot Kimi K3
The global AI industry is turning its attention back to China as the Chinese AI startup Moonshot has released its large-scale open-weight model, 'Kimi K3.'
If the emergence of 'DeepSeek' early last year served as an opportunity to confirm China's capability to develop low-cost, high-performance AI, the launch of Kimi K3 is being evaluated as a case demonstrating that China is entering the top tier of the relevant market, rapidly narrowing the AI technology gap with the United States despite cutting-edge semiconductor export restrictions.
Kimi K3 drew attention immediately after its release on July 17 due to its overwhelming model scale and performance.
According to the developer, Moonshot, Kimi K3 is one of the world's largest open-weight AI models, equipped with 2.8 trillion parameters.
Parameters are a type of internal variable that reflects the patterns and relationships of data during the AI model's training process and are considered a key indicator of a model's scale and complexity.
While direct comparison is difficult because closed-source, cutting-edge US models do not disclose such figures, it is considered the largest among open-weight models.
The company explained that K3 focuses on 'agentic AI' that autonomously performs complex tasks such as long-term coding, professional knowledge work, and reasoning, rather than simple Q&A.
In initial performance evaluations, it stood out particularly in the field of coding.
According to the Global Times, an English-language newspaper affiliated with the People's Daily, the official newspaper of the Communist Party of China, K3 scored 1,679 points in the web development task capability evaluation on the AI system evaluation platform Arena on the day of its release, surpassing Anthropic's 'Claude Fable 5' (1,631 points) to take first place.
AI performance analysis firm Artificial Analysis evaluated that K3 outperformed Anthropic's 'Opus 4.8' in some 'frontier benchmarks' (top-tier performance evaluation indicators).
Moonshot stated that K3's overall performance is still lower than Claude Fable 5 and OpenAI's GPT-5.6, but the market is paying attention to China's rapid technological pursuit.
Ion Stoica, a professor at UC Berkeley and founder of Arena, told Bloomberg, "I previously thought Chinese AI models were 6 to 9 months behind the top US models, but the current gap could be about 2 to 3 months."
Bloomberg analyzed that Kimi K3 could exert pressure on US AI companies to lower prices.
An executive at a US AI research lab, who requested anonymity, emphasized that while US companies will continue to differentiate themselves through continuous innovation, increased competition will require them to prove higher value not only in model performance but also in pricing, data integration, and user experience.
However, some point out that it is too early to place Kimi K3 on the same level as verified US AI models.
Unlike DeepSeek, which garnered global attention with claims of developing competitive models at a much lower cost than its US rivals, Kimi K3 is on the higher side in terms of price when compared to similar open-weight models.
The cost per task for Kimi K3 calculated by Artificial Analysis is $0.94 (approximately 1,391 won), which is cheaper compared to GPT-5.6 Sol ($1.04), Claude Fable 5 ($2.75), and Opus 4.8 ($1.80), but significantly higher than similar models such as GLM-5.2 ($0.32) and DeepSeek V4 Pro ($0.04).
Moonshot is preparing to raise large-scale funds as it pursues an IPO on the Hong Kong stock market within 6 months.
The company's valuation is expected to be assessed at over $30 billion (approximately 45 trillion won) in the current investment round.
As Kimi K3 received relatively favorable reviews in its early stages, Chinese state-run media emphasized that this model demonstrates China's innovation, efficiency, and openness.
The Global Times cited industry analysts as saying, "Despite US technology sanctions, Chinese AI companies are emerging as competitors equipped with both technical prowess and cost efficiency, moving beyond the role of simple followers."
Liu Gang, chief economist at the Chinese Institute for New Generation Artificial Intelligence Development, told the outlet, "Kimi K3 shows the speed and direction of China's technological innovation," and Chen Jing, vice president of the Institute of Technology and Strategy, argued, "Various models such as DeepSeek, Qwen (Alibaba), and GLM (Zhipu) are developing along different technological paths, reducing the risks associated with relying on a single technology approach."
(Photo: Yonhap News)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.