Chinese Open-Source AI Models Challenge Silicon Valley Approach

Chinese AI laboratories continue their offensive by releasing near-frontier open models. Z.ai released GLM 5.2 in June, Moonshot AI released Kimi K3 last week, and Alibaba released Qwen 3.8 this Monday. These models, especially K3, sparked immediate reactions in Silicon Valley and Washington. Venture capitalist and President Trump's AI advisor David Sacks called the performance of Moonshot's model "alarming." Commerce Secretary Scott Bessent suggested the possibility of imposing sanctions on Chinese AI companies. White House Office of Science and Technology Policy Director Michael Kratsios stated that the administration has "information that Moonshot AI distilled Fable from Anthropic to develop its K3 model," calling it "theft of proprietary American technology and undermining US research." Moonshot AI did not respond to a request for comment. Chinese models share common features: based on third-party benchmarks, they are nearly on par with the best Western models, optimized for agentic programming tasks, and are released or will be released with open weights. This mirrors the situation when DeepSeek R1 was released in January 2025, when DeepSeek's open approach challenged the dominance of closed models. Since then, Western labs have continued to move toward even greater closedness: Anthropic limited Mythos and Fable 5, OpenAI delayed GPT 5.6 at the White House's request. Meanwhile, Chinese startups and giants have doubled down on openness—their models can be downloaded, run locally, and modified. Arena AI platform ranks K3 as the best model for web development and fourth for agentic tasks, while Artificial Analysis places K3 third in its intelligence index. After the preview of K3 was released on July 16, demand was so high that Moonshot temporarily restricted new user registrations. Some are beginning to question whether it is worth paying for OpenAI or Anthropic. Independent researcher Nathan Lambert believes Anthropic exaggerated the risks. Chinese models have already found practical applications: after GPT-5.6 Sol hacked the Hugging Face system, the platform used GLM 5.2 to analyze the attack because other models refused to help due to built-in restrictions. Although Chinese models are cheaper per token, they may require more tokens to solve tasks, narrowing the cost difference. Former White House AI advisor Dean Ball, while praising K3, noted that the model is "very token-hungry" and questioned its actual cost-effectiveness. Nevertheless, such models undermine the fundamental assumption that infinite compute investment is necessary to create the best models.