Apple Trains Its Own AI Model for China With Alibaba's Help

Apple has trained a custom AI model for the Chinese market in partnership with Alibaba, according to three people familiar with the matter, a shift from its earlier strategy of relying on existing Chinese AI providers to bring generative AI to devices sold in the country (Reuters).

The model is built on Alibaba's Qwen architecture, Reuters reported. Apple had previously partnered with established Chinese AI companies rather than building its own model for the region. Training a custom model with Alibaba's support gives Apple tighter control over the on-device experience — that is, the AI processing that happens directly on the phone rather than in the cloud — while still navigating the regulatory constraints that govern AI in China.

US AI models that Apple uses in other markets, such as OpenAI's ChatGPT, are not available in China (The Verge). That gap has left Apple needing a China-specific approach to deliver Apple Intelligence features to its third-largest market.

Apple has officially registered its on-device generative AI service with China's cyberspace regulator, clearing a major regulatory hurdle. Chinese law requires AI models to be registered with and approved by the government before public release. A Reuters report in mid-July indicated that the registered Apple Intelligence service will incorporate capabilities from AI models developed by both Baidu and Alibaba (Reuters). The August reporting clarifies that Apple's custom model, built with Alibaba, is the vehicle for this rollout.

According to Reuters, Apple would become the first US company approved to offer a proprietary AI model in China. Apple Intelligence is expected to launch in the country in the coming months following an iOS update (The Verge).

The Alibaba partnership has been building for over a year. Alibaba's chairman confirmed on February 13, 2025 that the company would partner with Apple to support AI services on iPhones sold in China (Reuters). More recently, Apple said Mac users in China can connect to Alibaba's Qwen AI service, signaling that the integration extends beyond iPhone to the broader Apple ecosystem (Reuters).

Alibaba's Qwen model family has been a credible contender in the open-weights LLM space — meaning the model's parameters are publicly available for others to study and build on. In January 2025, Alibaba released Qwen 2.5, claiming it surpassed DeepSeek's V3 model (Reuters). Alibaba Cloud has also been active in the model ecosystem more broadly; it launched the first training and deployment solution for Meta's Llama 2 series in China in July 2023 (Reuters).

The regulatory architecture here is worth understanding clearly. China's Cyberspace Administration requires that any generative AI model available to the public undergo a registration and review process, covering content safety, data governance, and algorithmic transparency. For a foreign company, the path is narrower still: operating through a domestic partner and conforming to content moderation expectations that do not apply in Western markets. Apple's decision to train its own model with Alibaba, rather than simply reskin or call a domestic provider's API (the interface that lets one application use another's services), suggests an effort to keep inference and user data handling within Apple's control as much as the regulatory framework permits.

The distinction matters for Apple's privacy posture. In other markets, Apple Intelligence processes requests on-device where possible and uses Private Cloud Compute for more complex queries, with Apple asserting that user data is not stored on its servers. A custom model built with Alibaba's architecture could, in principle, preserve that on-device-first design philosophy in China. Whether Chinese regulators will permit the same data-handling guarantees that Apple advertises elsewhere is an open question that the rollout will answer in practice.

The competitive landscape adds pressure. Chinese smartphone manufacturers have been integrating generative AI into their devices at a rapid pace, and Apple's smartphone market share in China has faced headwinds. Delivering Apple Intelligence to Chinese users has been a deferred obligation since the feature launched in Western markets. The regulatory registration and the custom model together remove the two principal blockers: legal clearance and model availability.

The broader tension here is worth flagging. Apple has positioned privacy as a core differentiator, and building an AI model that complies with Chinese content regulations means operating within a framework that includes government oversight of model outputs. How Apple communicates the privacy characteristics of its China-specific model to users, and whether it discloses the extent of Alibaba's involvement in training or ongoing operation, will test the company's ability to maintain a consistent global privacy narrative across jurisdictions with fundamentally different expectations.