Staff at three Chinese AI labs that release open weights, Tencent, DeepSeek and Alibaba’s Qwen, named five distinct caveats on openness in public remarks reported between November 4 and November 8, by our count of the remarks as reported in our own coverage. Tencent’s chief AI scientist Yao Shunyu was asked whether Tencent would keep its Hunyuan model open if it became the world’s top-ranked model and gave no direct yes or no answer, Chinese financial outlet Yicai reported on November 8, as we reported.
The five caveats, by our count: from Tencent’s Yao, a harm threshold for cyber and bio abilities and a motive tied to not yet leading; from DeepSeek’s Gao Xi, amplified safety risk. From Alibaba Qwen’s Liu Dayiheng: an open model cannot be universally patched or recalled, and its safety alignment can be removed. Chinese models make up 89% of enterprise open-source model spend, Menlo Ventures said in a report we covered on November 3.
Business Pill · API vs Open Weights
A short Business Engineer explainer on renting a model through an API versus running open weights yourself. It teaches the general idea only and does not discuss these labs.
The key insight: By our count, staff at Tencent, DeepSeek and Alibaba’s Qwen named five distinct caveats on open weights in remarks reported between 4 and 8 November, while Tencent’s Yao Shunyu gave no direct yes or no on whether Hunyuan stays open if it leads, Yicai reported. Chinese models make up 89% of enterprise open-source model spend, Menlo Ventures said.
Tencent. Yao gave three motives for opening models, Yicai reported: winning developers when a model is not yet in the lead, sharing knowledge with the technical community, and, increasingly, safety. “The cyber abilities of [AI] models and soon [the] bio ability of models will reach a threshold where it’s actually possible to do harm”, Yao said, the South China Morning Post reported. Yicai reported that Yao expects open and closed models to coexist.
DeepSeek. Speaking on a panel at the World Internet Conference in Wuzhen on November 3, DeepSeek research engineer Gao Xi said: “Open source may indeed amplify certain AI-related safety risks. But keeping models closed does not eliminate safety problems either”, adding that colleagues are working on approaches “such as model tiering”, according to Geopolitechs’ translation of the session, as we reported in an update. The South China Morning Post reported the remarks on November 4 as rare comments from a DeepSeek representative.
Alibaba Qwen. On the same panel, Alibaba Qwen project head Liu Dayiheng said an open model, once widely downloaded, cannot be universally patched or recalled, and that its safety alignment can be removed by post-training or fine-tuning, per the same Geopolitechs account.
What each lab’s representative said in public, as reported in our coverage, 3-8 November 2026: Tencent’s Yao Shunyu (Yicai, South China Morning Post), DeepSeek’s Gao Xi (Geopolitechs’ translation of the 3 November World Internet Conference panel; South China Morning Post) and Alibaba Qwen’s Liu Dayiheng (Geopolitechs’ account of the same panel). Filled = stated in the remarks we report; open ring = not stated. Open-weights status from our pieces on Hunyuan, DeepSeek and Qwen. Our count: five distinct caveats.
Where the Open Models Are Used
Menlo Ventures reported that Chinese models grew to 89% of enterprise open-source model spend, led by Kimi (61%), DeepSeek (15%), GLM (11%) and Qwen (3%). Open-source models handle 19% of enterprise AI tokens but take 4% of spend, Menlo said, after the average open-source token fell to $0.13 per million.
Nous Research released Arete-27B on November 6, a post-trained version of Alibaba’s open-weights Qwen3.8-27B, as we reported. DeepSeek Harness, DeepSeek’s open-source agent app, had 1.54 million daily active users on November 4, its team said, as reported by IT Home and in our coverage.
As we read Yicai’s account, Yao’s first motive ties openness to position: the stated reason to open applies to models not yet in the leading position.
As we read the Wuzhen remarks, DeepSeek’s Gao Xi and Alibaba Qwen’s Liu Dayiheng described risks that are specific to weights once they are downloaded, while Gao added that keeping models closed does not eliminate safety problems either.
As we read Menlo’s figures alongside the remarks, the labs voicing caveats supply the open models that Western developers, such as Nous Research with its Qwen-based Arete-27B, build on.
Gao Xi, DeepSeek research engineer, per Geopolitechs’ translation
“Open source may indeed amplify certain AI-related safety risks”
Three Implications
FOR BUILDERS ON CHINESE WEIGHTS Menlo Ventures reported that Chinese models make up 89% of enterprise open-source model spend, led by Kimi at 61% and DeepSeek at 15%, and Nous Research built Arete-27B on Alibaba’s open-weights Qwen3.8-27B.
FOR THE LABS’ OWN PRODUCTS DeepSeek Harness, DeepSeek’s open-source agent app, had 1.54 million daily active users on 4 November, its team said, as reported by IT Home.
FOR THE QUESTION TO WATCH Yicai reported that Yao gave no direct yes or no on keeping Hunyuan open if it ranked first in the world, and that Yao expects open and closed models to coexist.
The essay says the winning pattern is to “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”, and that “Timing, execution, and competitive dynamics determine whether the strategy actually works.”
As we read it, Yao’s answer is a timing condition in those terms: Yao tied openness to not yet leading and named a capability threshold for harm, while Gao and Liu named risks specific to weights once released. The essay does not discuss these labs.
What Is Not Established
None of the three labs announced a change to how it releases weights in the remarks we report. Yicai reported that Yao gave no direct yes or no, which is neither a commitment to close nor to stay open. Gao Xi and Liu Dayiheng were not asked the Hunyuan question. Our earlier piece reports Liu’s remarks in paraphrase, per Geopolitechs’ account. The count of five caveats is ours and depends on how the remarks are grouped. Menlo’s shares describe enterprise spend in its survey, not all use of open models.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Staff at Tencent, DeepSeek and Alibaba’s Qwen named five caveats on open weights in remarks reported this week, by our count, while Yao gave no direct answer on whether Hunyuan stays open if it leads, Yicai reported. Chinese models make up 89% of enterprise open-source model spend, Menlo Ventures said.
A note on sourcing. This angle piece sets side by side remarks and figures already published in our own coverage this week, each sourced there: Yao Shunyu’s remarks as reported by Yicai and the South China Morning Post, DeepSeek’s Gao Xi’s Wuzhen panel remarks per Geopolitechs’ translation and the South China Morning Post, Alibaba Qwen’s Liu Dayiheng’s per Geopolitechs’ account, Menlo Ventures’ 2026 State of Enterprise AI report, Nous Research’s model card and IT Home’s report of DeepSeek Harness’s user numbers. The five-caveat count and the chart are our own grouping of those remarks. We did not contact the labs.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
