China's open AI models rival Silicon Valley, fueling US concern

On Monday, Alibaba released Qwen 3.8 with open weights—the latest in a wave of capable open-source models from Chinese AI labs. Earlier in July, Moonshot AI released Kimi K3, which has been widely hailed as the strongest of the new releases and is now ranked fourth in agentic tasks on Arena AI and third on Artificial Analysis's intelligence index, just behind Anthropic's Fable and Opus 4.8 and OpenAI's GPT 5.6. Z.ai had released GLM 5.2 in June. All three models perform nearly as well as leading Western models on third-party benchmarks and are optimized for agentic coding tasks—the year's hottest area in AI.

The releases have triggered immediate concern in Washington and Silicon Valley. David Sacks, a venture capitalist and AI adviser to President Donald Trump, called Moonshot's K3 performance "concerning." Commerce Secretary Scott Bessent suggested the US might impose sanctions on Chinese AI companies. On Wednesday, Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged that "Moonshot AI distilled Anthropic's Fable for the development of its K3 model," calling this "stealing proprietary US technology and undermining American research" and "unacceptable." Moonshot AI did not immediately respond to the accusation.

The models' release strategy—open weights, free access, local deployment—stands in sharp contrast to the path Western labs have taken over the past year. Anthropic stated for months that its Mythos model was so capable at hacking that only approved collaborators could use it. When finally released more widely, the White House issued broad export controls, forcing Anthropic to temporarily take Mythos and its less capable sister model, Fable 5, offline. OpenAI similarly delayed releasing GPT 5.6 after receiving a request from the White House. In China, by contrast, startups and tech giants have doubled down on open source: anyone with sufficient hardware can download open-weight models, run them locally, customize them, and enjoy freedoms that OpenAI and Anthropic would never permit.

Chinese labs have adopted this strategy partly out of necessity. As newer, smaller players in the AI field, they can attract users, collaborators, and attention by making models free and open—and by doing so, they compete in a separate lane from the deep-pocketed giants. Alibaba had faced rumors earlier this year that it might shift to closed-source after reorganizing its AI teams, but the Monday release of Qwen 3.8 with open weights signaled no such pivot. The demand for K3 has been so overwhelming that Moonshot temporarily restricted new user sign-ups shortly after releasing a preview on July 16.

The practical impact is already visible. Nathan Lambert, an independent AI researcher, noted that even weeks after GLM 5.2's release, AI researchers in the Bay Area are using it as a core part of their workflow, and K3 "will only do more, especially in areas like cybersecurity, where Mythos, Fable, and GPT 5.6 are effectively unusable." Hugging Face, hit by a cyberattack in which OpenAI's GPT-5.6 Sol model hacked into its production systems, resorted to using GLM 5.2 to analyze the breach because other frontier models' built-in safety guardrails prevented them from helping. Dean Ball, a former White House AI adviser who recently joined OpenAI as head of strategic futures, acknowledged K3 as "a very good model" while noting it appeared "very token-hungry," making the cost advantage smaller than raw per-token pricing suggests. Yet Ball also observed that open-weight models fundamentally "deter further AI capex"—challenging the assumption that AI labs need unlimited funding to scale compute and build better models. Rui Ma, founder of the research firm Tech Buzz China, summed up the sentiment: the success of K3 was "only made possible by the poor comms and decisions from [Silicon Valley] labs in the past year."

The release of Moonshot AI's K3 and other open-source Chinese models represents a significant strategic divergence in how American and Chinese AI labs are approaching the market. Since DeepSeek's R1 model shocked the world in January 2025, Western labs have paradoxically moved in the opposite direction—toward more restriction. Anthropic's Mythos was withheld for months as too dangerous, then temporarily taken offline after White House export controls forced the company's hand. OpenAI delayed GPT 5.6 following a separate White House request. These moves have created an opening for Chinese startups, which have chosen to release capable open-weight models freely, allowing any developer with sufficient hardware to download, customize, and deploy them locally.

The White House has responded with alarm, with Michael Kratsios alleging that Moonshot distilled Anthropic's Fable for K3's development—a claim Moonshot has not addressed. Commerce Secretary Scott Bessent suggested possible sanctions. Yet the response comes as Chinese models have already begun displacing Western ones in real-world use. Hugging Face, an open-source AI platform, turned to GLM 5.2 to analyze a cyberattack because frontier models' safety guardrails made them unsuitable. Researchers in the Bay Area report continuing to use GLM 5.2 for core workflows. This practical utility, combined with the models' availability and near-parity performance on benchmarks, is reshaping developer choices in ways that American export controls and access restrictions appear unable to prevent.

No comments yet. Be the first to share your thoughts!