China’s Moonshot AI unveils Kimi K3, closing gap with U.S. AI leaders

China's Moonshot AI unveiled Kimi K3, a 2.8 trillion-parameter AI model designed for coding, research, and multimodal tasks, narrowing the gap with leading U.S. AI systems. The model uses a mixture-of-experts architecture with 896 components, achieving 2.5 times greater scaling efficiency than its predecessor, and is available through Kimi's platform and API. Moonshot plans to release the full model weights by July 27.

Moonshot AI has unveiled Kimi K3, a 2.8 trillion-parameter artificial intelligence model designed for extended coding, research and multimodal tasks, bringing the Beijing startup closer to leading U.S. AI systems. In a technical blog published by the company, Moonshot described K3 as its most capable model to date, combining native image and video understanding with a context window of one million tokens. K3 is available through Kimi’s web platform, workplace tools, coding agent and developer API. Moonshot said it plans to publish the full model weights by July 27, with a technical report detailing its architecture, training and evaluations to follow. Kimi K3 packs 2.8 trillion parameters into a sparse expert design Moonshot called K3 the first open model to reach 2.8 trillion parameters, placing it in what the company describes as the three-trillion-parameter class. The model uses a mixture-of-experts architecture containing 896 specialist components, although only 16 are activated for each task. Moonshot said that design, combined with its Kimi Delta Attention and Attention Residuals technologies, delivers about 2.5 times greater scaling efficiency than Kimi K2. Despite the efficiency gains, operating K3 independently will require substantial infrastructure. Moonshot recommends systems containing at least 64 connected accelerators, highlighting the computing burden involved in deploying models of its size. Kimi K3 builds software, studies papers and generates thousands of code lines Moonshot positioned K3 as a long-horizon agent capable of navigating large software repositories, operating terminal tools and completing complicated engineering projects with limited human supervision. In company demonstrations, K3 built a compact GPU compiler and designed, optimized and simulated a small AI chip during a 48-hour autonomous run. Moonshot also said the model recreated a computational astrophysics study in about two hours after reviewing more than 20 research papers and generating more than 3,000 lines of Python code. The company also highlighted K3’s ability to create interactive research reports, edit video and turn images or concepts into playable digital experiences, saying its internal tests placed the model ahead of rivals across several coding, knowledge-work and visual benchmarks. https://thecoinheadlines.com/tech-and-ai/29-countries-including-russia-join-chinas-world-ai-cooperation-organization/article-26444/ Moonshot also acknowledged that K3 still trails the strongest proprietary models in overall performance and user experience, leaving wider testing to determine whether its enormous scale translates into a consistent advantage.