Version 7

Moonshot AI's Kimi K3 (2.8 trillion parameters, open weights) and Alibaba's Qwen3.8-Max (2.4 trillion parameters, API preview only) are the leading Chinese open-weight models, assessed by analysts at 3–7 months behind closed frontier systems [2][3][4]. A Bloomberg investigation reports that Moonshot trains on roughly 20,000 Nvidia chips — possibly H200s — accessed through a confidential deal with Alibaba, raising direct questions for US export control enforcement [13]. Third-party benchmarks find Qwen3.8-Max outperforming Claude Fable 5 on 3D physics tasks at roughly 7x lower cost [8], and 1-bit quantization has compressed K3 to 590GB for local deployment on four B200 GPUs [16]. DeepSeek's V4-Flash-0731 (304B parameters) adds a third competitive Chinese release alongside the two flagship models [9].

The Moonshot-Alibaba chip arrangement moves the export control debate from theoretical to potentially concrete: if the supplied chips are H200s, which are export-restricted to China, that is a live enforcement question rather than a policy proposal [13]. At the same time, Qwen3.8-Max's cost-performance results and K3's quantized local deployability continue to show the open-to-closed gap closing in specific task domains.

Moonshot AI's Kimi K3 (2.8 trillion parameters, mixture-of-experts) was released as open weights in July 2026 and is available on Hugging Face at 1.56TB [1]. Analyst estimates place it within 3–7 months of closed frontier systems: Lambert at 3–5 months, Mowshowitz at 4–6 months, and the UK AI Security Institute at 4–7 months for cybersecurity tasks specifically [2][3][4]. Alibaba's Qwen3.8-Max (2.4 trillion parameters) was previewed at the same time with self-reported benchmarks claiming it trails only Claude Fable 5; it remains an API preview only, despite a commitment to open-weight release made following Xi Jinping's July 20 WAIC keynote [5][6]. Third-party evaluations have begun: SemiAnalysis published performance tests of the preview [7], and a coding benchmark found Qwen3.8-Max outperformed Claude Fable 5 on 3D physics scene generation at roughly 7x lower cost — $0.28 per task versus Fable 5's $1.93, with Qwen demonstrating stronger geometric accuracy and collision handling [8]. DeepSeek's V4-Flash-0731 (304 billion parameters, released July 31) adds a third Chinese model at $0.14/$0.27 per million tokens, ranked by Artificial Analysis ahead of MiniMax M3 on its Intelligence Index [9].

A substantive disagreement has developed over what drives Chinese model performance. Nathan Lambert argues the gains are genuine innovation: RL training — where real capability advances occur — requires millions of rollouts prohibitively expensive to run through closed-model APIs, making distillation a limited contributor [10]. He explicitly calls Ben Thompson's claim that distillation scales with RL factually incorrect and potentially misleading to policymakers [10]. Zvi Mowshowitz holds that K3's benchmark scores are systematically overstated due to 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 [3]. Semafor reports K3 underperforms top frontier models on cybersecurity vulnerability detection specifically, suggesting capability is uneven across domains [11]. Florian Brand, writing in Interconnects, argues that earlier predictions of open-weight ecosystem consolidation have failed to materialize — more companies than ever are training and releasing frontier-scale models — and that commercial viability is demonstrated by Thinking Machines, whose open-model fine-tuning service generates hundreds of millions in annual revenue [12].

A Bloomberg investigation adds a concrete dimension to the export control debate. Moonshot AI reportedly trains its Kimi models on roughly 20,000 Nvidia chips accessed through a confidential deal with Alibaba, one of Moonshot's largest investors. Multiple sources identified the chips as H200s, which are export-restricted to China under US rules. Alibaba denied the H200 identification but did not deny supplying 20,000 chips and refused to disclose what they are [13]. The Trump administration was separately reviving efforts to restrict Chinese AI models after K3's release [14], while Congress introduced bipartisan kill-switch legislation assessed as ineffective against sophisticated systems [11]. Lambert introduced a concrete cost to a ban: a Chinese open model was the only tool available to analyze an OpenAI breach because US frontier models had guardrails blocking the analysis, meaning restrictions would directly disadvantage American cybersecurity defenders [10]. An Anthropic cybersecurity incident — a configuration mistake that left a supposedly isolated test environment open to the internet, allowing Claude to reach real companies — has been cited as additional relevant context [15]. Brand notes that K3's commercial license, requiring MaaS businesses over $20 million in annual revenue to negotiate separate agreements with Moonshot, could create regulatory leverage for US government action against American entities using the model [12].

1-bit quantization has compressed Kimi K3 from 1.56TB to 590GB — a 62% reduction — while retaining 78.7% agreement with the original model and its full 1-million-token context window [16]. Increasing to 2.1-bit raises agreement to 87.2% at 769GB. In a practical HTML 3D physics test, the locally-run 1-bit K3 was the only model among four tested — including the Kimi K3 API, Claude Opus 5, and GPT 5.6 — to build a working winch mechanism, and running it locally on four B200 GPUs costs nothing in API fees versus $0.30–0.77 for cloud alternatives [16]. A public letter on Open Weights and American AI Leadership, signed by nearly all major AI figures with one notable exception, reflects a consolidating practitioner consensus that open-weight development should continue [17].