China's Moonshot AI chases 'DeepSeek moment' with much-hyped model

A Chinese artificial intelligence startup, Moonshot AI, unveiled a new large language model named Kimi K3 on July 17, sparking discussions about the nation’s rising technological capabilities. The model, described as the world’s first open-source model of its scale, has drawn comparisons to a landmark 2025 release from another Chinese firm, DeepSeek, which challenged perceptions of U.S. dominance in the field. According to reports, Kimi K3 achieved high rankings in evaluations conducted by a research group at the University of California, Berkeley, and has already begun drawing attention from both industry experts and investors. The release of Kimi K3 comes amid a broader trend of growth among Chinese AI companies, who are increasingly gaining traction globally due to the affordability of their products and the availability of customizable source code. This model, with approximately 2.8 trillion internal variables, known as parameters, is designed to process complex queries more effectively than previous iterations. While Moonshot AI claims that Kimi K3 demonstrates “frontier-level” performance in certain areas, it acknowledges that it still lags behind the most advanced proprietary models developed by U.S.-based firms like Anthropic and OpenAI. Experts have offered mixed assessments regarding the significance of Kimi K3. Ethan Mollick, a professor at the University of Pennsylvania and prominent figure in AI research, noted that while the model performs well in specific domains, it still struggles with creative tasks such as writing compelling fiction. He stated that generating a good murder mystery remains a challenge for all current AI models. Meanwhile, tech writer and investor Kevin Xu suggested that the market response to Kimi K3 might mirror the impact of DeepSeek’s 2025 release, which had previously shaken confidence in U.S. leadership in AI. Despite these reservations, the performance of Kimi K3 has been impressive enough to generate interest within the academic community. Hussein Abbass, a computing professor at UNSW Canberra, highlighted that while the model shows promise in coding-related tasks, its competitiveness across a wide array of functions typical of foundation models remains uncertain. He emphasized that AI performance depends not only on the model itself but also on the supporting infrastructure, including hardware and data centers. In practical tests, Kimi K3 produced results comparable to those of other leading chatbots when asked to generate ideas for technology journalism. It ranked ninth globally for text queries according to the Arena platform, which evaluates AI models based on their ability to answer questions accurately and coherently. This achievement positions Kimi K3 as a notable contender in the rapidly evolving landscape of AI technologies. The release of Kimi K3 coincided with a major technology conference in Shanghai, where President Xi Jinping addressed the importance of international collaboration in governing AI advancements. This context underscores the geopolitical dimensions of AI development, particularly given the ongoing restrictions imposed by the United States on exporting high-performance microchips used in training and running AI systems. These restrictions aim to prevent China from catching up too quickly in this critical area of technological competition. Recent developments suggest that regulatory considerations may influence the pace at which top-tier AI models are made publicly available in the U.S. The Trump administration has reportedly delayed the release of advanced models from Anthropic and OpenAI, citing security concerns related to potential vulnerabilities in online systems. As a result, there is speculation about whether these U.S. firms might accelerate their release schedules in response to the growing capabilities of Chinese counterparts like Moonshot AI.