China's Moonshot AI Unveils High-Performance, Low-Cost AI Model 'K3' to Rival US Tech Firms

Comparable performance at a third of the price. Chinese AI startup Moonshot AI has unveiled a new large language model, 'Kimi K3,' that rivals top-tier U.S. models. While K3 boasts similar performance to 'Fable 5' by American AI firm Anthropic, its usage fee is significantly lower at $15 per 1 million output tokens. Tokens are the units of information processed and generated by large language models. The model adopts a Mixture-of-Experts (MoE) architecture based on approximately 2.8 trillion parameters. Moonshot AI is offering K3 via an open-weight strategy. Open weights involves releasing the model's core data, allowing developers to directly modify and train the model themselves.

The release of K3 triggered reactions across global markets. On the 17th, TSMC shares fell 7%, and SoftBank stock also recorded a 9.0% drop. z.ai shares plummeted significantly, and U.S. Nasdaq and major AI-related stocks like NVIDIA and Meta also moved in tandem.

Experts believe the AI technology gap between the U.S. and China is narrowing faster than expected. Some suggest the gap, previously estimated at up to six months, could shrink even further. As China's AI ecosystem grows rapidly, shifts are appearing in the global competitive landscape. Currently, China's AI industry sees fierce competition ranging from giants like Alibaba and ByteDance to startups like DeepSeek and Minimax. Meituan recently released the 'LongCat 2.0' model, trained exclusively on Chinese semiconductors rather than U.S. chips. DeepSeek unveiled its 'V4' model back in April, and z.AI released its 'GLM-5.2' model in mid-June. Based on the OpenRoute weekly rankings, models from Chinese companies Tencent, Xiaomi, DeepSeek, Minimax, and z.ai claimed the top five spots.