China’s open-source AI strategy has triggered one of the sharpest policy rows Silicon Valley has seen in years, after Moonshot AI released Kimi K3 on 16 July 2026, a model with 2.8 trillion parameters that its developer describes as the world’s first open 3T-class system.
How Kimi K3 Ignited the Debate
Kimi K3 cleared several high-profile benchmarks at a cost well below comparable US models. On GDPval-AA v2, a benchmark measuring real-world tasks across 44 occupations and 9 major industries, VentureBeat reported the model scored 1,687, placing it third overall behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), and ahead of Claude Opus 4.8 (1,600), citing public leaderboard data and a private evaluation by analytics firm Artificial Analysis.
The model’s API pricing, listed on BenchLM, is $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million. The context window runs to 1.05 million tokens.
Moonshot AI claims roughly a 2.5x improvement in scaling efficiency over its predecessor Kimi K2, attributing the gains to two architectural changes: Kimi Delta Attention, a hybrid linear attention scheme, and Attention Residuals, which alter how information moves between layers, according to Tom’s Hardware. Bank of America analysts led by Alex Liu told CNBC: ‘Despite persistent hardware/compute capacity constraints in China, K3 demonstrates that pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models.’
The release did not come from nowhere. After DeepSeek’s R1 launch in January 2025 pushed Moonshot’s Kimi app from third to seventh in monthly active users in China, the company pivoted sharply toward open-source, beginning with Kimi K2 in July 2025 and continuing with K2.5 in January 2026, according to VentureBeat. Its Mooncake infrastructure project, which won the Best Paper award at FAST 2025, was designed to make inference at the scale of a 2.8-trillion-parameter model practically viable.
A Strategy Divide With Global Stakes
China’s open-source AI strategy produces models that anyone can download, modify, and deploy. The contrast with OpenAI and Anthropic’s closed approach sits at the heart of a weekend argument that spilled well beyond the usual tech circles.
Dean Ball, OpenAI’s head of strategy and a former senior adviser on AI to President Donald Trump, posted a lengthy reaction on X. ‘I am personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks,’ he wrote. He warned that open-weight models would lead to ‘AI communism’ and called them ‘decelerationist’ because they ‘deter AI capex.’
Ball went further, suggesting the Trump administration’s best move would be to ‘create large amounts of regulatory risk around the use of open-weight Chinese models.’ He outlined one mechanism in detail: directing agencies to issue soft guidance, citing as an example ‘A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models,’ to ensure that regulated enterprises back away from Chinese open-weight products, according to Cybernews. Ball later clarified the remarks were a prediction rather than a recommendation, and that he supports open-source up to the point where AI becomes too dangerous.
The suggestion drew immediate fire. David Sacks, who co-chairs the President’s Council of Advisors on Science and Technology (PCAST) alongside Michael Kratsios following the council’s establishment by executive order on 23 January 2025, and whose White House appointment was announced in March 2026, called the ‘weaponization of regulatory uncertainty’ ‘completely unacceptable.’
‘We are at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open source competition,’ Sacks wrote, naming OpenAI and Anthropic as that duopoly. ‘They have laid their cards on the table. It is time for the rest of Silicon Valley, the vast majority that still values open competition, to do the same.’
Venture capitalist Chamath Palihapitiya, Sacks’s co-host on the All-In podcast, was equally direct: ‘The future is open source. We need to embrace it and get on with it.’ Software entrepreneur Suhail Doshi argued that lobbying against open-weight models on distillation grounds was ‘total BS,’ contending that American AI labs trained on ‘humanity’s data and didn’t pay a cent.’
Not everyone framed the picture as a Chinese advance. A Citrini Research analyst known on X as Jukan pushed back on the takeover narrative, arguing that DeepSeek’s efficiency advantages stem from its proprietary operations rather than open-source code alone. ‘Chinese companies may lack sufficient compute capacity to serve all the inference demand themselves, but they are not selling at a loss or failing to recoup their training costs,’ Jukan wrote.
Ball’s soft-law proposal effectively defines the next test. If the Trump administration moves to formalise regulatory doubt around Chinese open-weight models, it will force every regulated US enterprise to choose a side, well before Congress has written a single line of AI law.