Does China’s latest AI model finally equal US rivals? What scientists think

Last week, Beijing-based Moonshot AI introduced its latest large language model, Kimi K3, marking a notable advancement in China's artificial intelligence landscape. According to internal testing conducted by the company, K3 demonstrates the ability to match or surpass U.S.-based models in specific tasks such as coding and spreadsheet manipulation. This achievement has sparked considerable interest among scientists and industry experts, with some suggesting it represents a pivotal moment in the global race for AI supremacy. The launch of Kimi K3 occurred on 16 July, coinciding with the upcoming 2026 Artificial Intelligence World Conference in Shanghai. During this event, Chinese President Xi Jinping announced the initiation of a global initiative aimed at establishing regulations that ensure AI serves the public good. He emphasized the importance of a people-centered approach to AI development, advocating for shared prosperity and collective security through technological progress. Mehwish Nasim, an AI researcher at the University of Western Australia, noted that the timing of K3's release is strategically aligned with recent advancements from U.S. competitors. Just weeks prior, Anthropic had launched its Claude Fable 5, and shortly thereafter, OpenAI unveiled its GPT-5.6 model. Nasim highlighted that China's move signals a broader ambition to influence both the creation and governance of the international AI ecosystem. Kimi K3 is designed as an open-weight model, similar to its predecessors Kimi K2 and DeepSeek. This classification means that the model's core components are publicly accessible, allowing researchers to download and modify them freely, although detailed training data remains confidential. Access to the model is facilitated through an application programming interface, offering a more affordable alternative to proprietary models managed by entities like OpenAI, Google, and Anthropic. Moonshot plans to release the model's parameters on 27 July, following a period of high demand that led the company to temporarily halt new sign-ups due to system limitations. Rahul Shome, a robotics and AI researcher at the Australian National University, expressed enthusiasm over the potential impact of open-weight models, noting a growing consumer and academic preference for transparency and accessibility in AI technology. K3, with its impressive scale of 2.8 trillion parameters and a working memory capable of handling up to one million tokens, stands out as the largest open-weight model to date. Niusha Shafiabady, a computational intelligence researcher at Australian Catholic University, pointed out that K3's extensive working memory could significantly enhance accuracy by enabling the model to retain vast amounts of textual information, thus reducing the risk of generating misleading content known as hallucinations. She intends to personally evaluate the model's effectiveness by requesting it to summarize scientific research findings. Aaron Snoswell, an AI accountability researcher at the Queensland University of Technology Generative AI Lab, observed that while open-weight models have traditionally underperformed relative to their U.S. counterparts, the current gap appears to be narrower than ever. This shift reflects evolving perceptions among investors and users regarding the value proposition of different AI models, influenced by the substantial financial investments required to train leading-edge models in the United States versus the more restrained budgets of Chinese firms.