Version 3

Moonshot AI's Kimi K3 (2.8 trillion parameters, open weights) and Alibaba's Qwen3.8-Max (2.4 trillion parameters, open-weight release committed) have brought Chinese models to within 3–7 months of closed frontier systems by most analyst estimates [2][5][6]. Nathan Lambert's July 22 follow-up places Kimi K3 at roughly GPT-5.4–5.5 on coding tasks and argues Chinese gains reflect genuine innovation rather than distillation-driven copying, directly challenging Ben Thompson's framing [1]. The Trump administration is reportedly pursuing a ban on Chinese AI models, but the debate has expanded: Lambert argues such a ban would hurt US cybersecurity defenders who rely on these models for analysis tasks closed US models refuse to perform [8][1].

Chinese labs are now releasing near-frontier open-weight models on 1–2 month cycles [1], compressing the time before high-capability AI is broadly accessible. The US policy options are doubly constrained: open-weight distribution makes a ban structurally difficult to enforce, and restricting these models may disadvantage American defenders relative to attackers globally who face no such restriction.

Moonshot AI's Kimi K3, a 2.8 trillion-parameter mixture-of-experts model released as open weights in July 2026, is assessed by multiple analysts as the strongest open-weight model to date — performing at approximately GPT-5.4 to 5.5 level on coding tasks with strong agentic research capabilities [1][2]. Alibaba's Qwen3.8-Max (2.4 trillion parameters) was previewed with self-reported benchmarks claiming it trails only Claude Fable 5, with open-weight release committed following Xi Jinping's July 20 WAIC keynote publicly endorsing open-source AI [3][4]. Multiple analysts place the open-to-closed capability gap at 3–7 months, down from 6–10 months in prior periods [2][5][6]. Independent testing placed Kimi K3 fourth among all evaluated models [7].

A substantive debate has formed around distillation's role in Chinese model gains. Nathan Lambert argues Chinese labs' performance is genuine innovation: RL training — where he contends real capability gains occur — requires millions of rollouts that would be prohibitively expensive and slow to run through closed-model APIs, meaning distillation from those APIs has limited practical impact on model quality [1]. He characterizes these as models people compare favorably on internal training benchmarks, not distillation artifacts, and explicitly calls Ben Thompson's argument — that distillation becomes more impactful as RL scales — factually incorrect and potentially misleading to policymakers [1]. Zvi Mowshowitz holds a different position: Kimi K3 benchmarks are systematically inflated by max-effort evaluation settings, jagged capability profiles, and some Claude distillation, though he concedes the model is real and roughly four to six months behind the closed frontier [6].

The Trump administration is reportedly reviving a push to ban Chinese AI models following Kimi K3's launch, citing cybersecurity concerns, with officials acknowledging that open-weight distribution makes a comprehensive ban nearly unenforceable [8]. Lambert introduces a concrete counterargument: Hugging Face was able to analyze the OpenAI breach only by using a Chinese open model, because US frontier models had guardrails blocking the analysis — meaning a ban would directly disadvantage American cybersecurity defenders relative to global attackers who face no equivalent restriction [1]. This extends the enforcement problem from structural (open weights cannot be recalled) to strategic (restricting these tools harms US security). Ben Thompson's earlier proposal — legislation making AI training data collection explicit fair use and barring distillation restrictions for US companies — represents an offensive alternative posture, while Demis Hassabis proposes a FINRA-style pre-release standards body [4][5].

Chinese labs' cost efficiency advantage stems from focused engineering teams with no distracting side projects, lower operational costs, and increasing access to domestic compute including Huawei Ascend chips [1]. The open-weight ecosystem has professionalized to the point where Chinese labs now release models on 1–2 month cycles matching closed lab iteration speeds [1]. Nathan Lambert's April 2026 structural prediction — that RL-era agentic deployment data would be the first domain where closed labs hold a durable edge, and that Chinese open-weight labs would face funding difficulties by late 2026 [9] — remains unresolved: K3 and Qwen3.8-Max are the first major tests of that hypothesis, and Lambert's own July 22 analysis concedes K3 exceeded expectations on closing the capability gap.