Open-Weight AI Models from China and Their Global Reach — 2026-09-24
Alibaba previewed the Qwen 4 lineup at Apsara 2026 while shipping Qwen-Image-2.1 as open weights, prompting questions about whether Qwen 4's promised 27B tier will actually be released. DeepSeek told investors it is training a 2-trillion-parameter model with plans for an 8-trillion model, and Beijing reportedly opened a data security probe into DeepSeek and Moonshot AI. Meanwhile, US intelligence warnings about Chinese model distillation and Qwen's discovery in the US Federal Register are sharpening the policy debate.
Open-Weight AI Models from China and Their Global Reach — 2026-09-24
Top developments
Alibaba previews Qwen 4 Max lineup — but no weights yet
At Apsara 2026 on September 22, Alibaba previewed Qwen 4 Max, Qwen 4 Flash, Qwen 4 Plus and Qwen 4 27B — without publishing pricing, weights, context lengths or benchmark scores for any of them. Commentary in Chinese-language media notes Moonshot released Kimi K3's 2.8T weights just 11 days after launch, while Alibaba's Qwen 4 27B tier "was only a word at the launch event," fueling skepticism about whether open weights will actually follow. The gap between promised and delivered open weights is becoming a key test of Alibaba's open-model credibility.
Qwen-Image-2.1 ships as a 7B open-weight image model — and tops image arenas
On September 20, Alibaba open-sourced Qwen-Image-2.1, a 7B-parameter unified image generator and editor with native transparency support, released under a research-only licence alongside two 9B prompt rewriters. Japanese analysis reports the model now holds the highest scores on two of Image Arena's open-model boards, though the licence restricts commercial use. The release redefines expectations for what a 7B image model can do.
DeepSeek eyes 2T and 8T parameter models
DeepSeek CEO Liang Wenfeng told investors the company is currently training a 2-trillion-parameter model, with plans for an 8-trillion-parameter model afterward; the current flagship V4-Pro has 1.6T total parameters. For scale: Moonshot's Kimi K3, at 2.8T parameters released under open weights, remains the largest open-weight model published to date. The roadmap signals DeepSeek intends to compete at the absolute frontier of model scale.
Beijing reportedly probes DeepSeek and Moonshot over data security
Shares of Chinese AI model developers fell after reports on September 23 that regulators opened a data security probe into DeepSeek and Moonshot AI, following Anthropic's report accusing Chinese firms of model distillation. China's Cyberspace Administration reportedly first summoned representatives from all seven Chinese companies named in Anthropic's report before narrowing its focus to DeepSeek and Moonshot. The scrutiny comes just ahead of Trump-Xi AI talks.
Qwen found in US Federal Register; intel agencies warn on distillation
Days after the FBI publicly accused Alibaba of maliciously copying Anthropic's technology, it emerged that the US Federal Register website had used Alibaba's Qwen as a search tool, drawing bipartisan concern. Separately, three US intelligence agencies warned that DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI have used distillation tactics to copy US frontier model capabilities. The episode highlights the awkward reality that Chinese open-weight models are already embedded in Western public infrastructure even as restrictions are debated.
Local view
Korea: Tokenpost reports that Qwen-family derivative models on Hugging Face have reached 151,448, arguing that AI competitiveness should now be measured by developer adoption and ecosystem spread, not just model performance. Korea's Personal Information Protection Commission has issued recommendations to DeepSeek following findings that it responded in Chinese to a Korean user's privacy-related query, and flagged deficient privacy policies. Korean-language AI commentary is also examining how Qwen's licence terms and security responses affect self-hosting decisions.
Japan: ITmedia analysis asks why Japanese companies rely on Qwen and what an exit strategy would look like if the US government banned it — a sign that sovereign-AI risk around Chinese open weights has entered mainstream Japanese business discourse. PC Watch covered the pre-announcement of Qwen-Image 2.1's open-weight release for Japanese developers.
China/Chinese-language: Zhihu's model tracker lists updated open-weight agent releases including Xiaomi's MiMo-2.6-Flash and MiMo-2.7-Pro for the week of September 21–25.
Context & numbers
- Qwen derivatives on Hugging Face: 151,448 models counted as of mid-September, the largest open-model ecosystem in existence.
- Hugging Face Hub download data as of September 18, 2026 shows Qwen families dominating the most-downloaded open models, with Kimi K3, Florence-2 and Chandra OCR each single repositories under 3.4 million downloads.
- Kimi K3: 2.8T parameters, released as open weights July 27, 2026 — still the largest published open-weight model.
- Artificial Analysis independently scores Qwen3.8-27B at 52 on its Intelligence Index, level with GLM-5.2 and DeepSeek V4 Flash.
On the radar
- Kimi K3.1 leak: Kimi subscriptions reopened September 18, and a tipster claims Kimi's next model is in post-training, targeting an October release — unconfirmed rumor.
- Xiaomi MiMo-V2.6-Pro: Xiaomi's public RL dashboard reads 72.57 on DeepSWE, 1.4 points off the top, but no release date has been announced.
- Quiet release to watch: inclusionAI's MIT-licensed 6B Ming-Image-0.1-Design shipped September 17 with no announcement, no API endpoints, and a 404 quickstart page.
- Trump-Xi AI talks: the reported DeepSeek/Moonshot probe and US intel warnings set the backdrop for US-China AI discussions; watch for licensing or hosting restrictions emerging from these channels.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
