Chinese AI firms keep new models closed as cloud demand rises

Chinese AI firms including Alibaba Cloud and Beijing-based Zhipu AI are keeping some new models closed and offering them mainly through cloud services as bigger systems get harder to run on local hardware.

Alibaba this week launched three proprietary Qwen models available only via its cloud platform or chatbot site, including Qwen3.6-Plus for coding and the multimodal Qwen3.5-Omni, after open-sourcing the previous-generation Qwen3-Omni model released in September.

Alibaba Cloud said it chose not to open-source the latest Omni model because the Omni line has seen fewer downloads on Hugging Face, and it has also released newer Wan image and video generation models as proprietary software after US researchers linked earlier open versions to the spread of NSFW content.

Zhipu AI, known as Z.ai internationally, released GLM-5-Turbo last month, and its CEO said more customers are shifting from local deployments to cloud APIs, while closed-model products like Kuaishou’s Kling AI reported 340 million yuan in Q4 revenue.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Alibaba keeps some AI models closed to help pay for growth outside core commerce

Alibaba’s results split into two lines, an older e-commerce operation under strain and a fast-growing cloud plus AI unit 1.
Customer management revenue (CMR) rose 1%, while the China E-commerce Group’s adjusted EBITA (earnings before interest, taxes, and amortization) dropped 43% in a recent quarter 1.
Cloud Intelligence Group revenue climbed 36%. AI-related products kept triple-digit growth for ten consecutive quarters 1.
Keeping some models proprietary and selling them mainly through cloud services supports the goal to exceed $100 billion in combined cloud plus AI external revenue within five years, as commerce profitability and cash generation face pressure 2.

China rules and compliance make open-source generative AI harder to run

China’s draft generative AI measures would hold providers responsible for content produced through application programming interfaces (APIs), though enforcement can differ by regime and final rules 3.
A cybersecurity standards proposal from TC260 (a Chinese technical committee that develops cybersecurity standards) lays out training data sampling. It calls for blacklisting a corpus if over 5% of sampled data counts as “illegal and negative information,” though it remains a proposed standard rather than a law 4.
China’s generative AI governance framework requires action after illegal content is found. Steps include stopping generation, removing content, and optimizing models, plus filings or security assessments in certain cases 5.
A closed model offered only through the cloud lets companies run monitoring and set access controls. It also speeds remediation compared with open-source models that third parties can modify outside the provider’s environment 3.