Apple Intelligence arrives in China with Alibaba's Qwen: iOS 27.3 in December, Siri without Siri AI

Apple Intelligence Arrives in China with Alibaba's Qwen: iOS 27.3 in December, Siri Without Siri AI

🔎 To enter China, Apple agreed to unplug its own brain

More than two years after its American launch (October 2024, iOS 18.1) and twenty months after its arrival in Europe (April 2025, iOS 18.4), Apple Intelligence is crossing its last great frontier: mainland China. And the way it's doing it says a lot about the state of the world. To satisfy local regulations, Apple isn't shipping its own models, nor those of OpenAI or Google — but Qwen, Alibaba's LLM.

The timeline is of military precision: developer beta on November 4, public rollout with iOS 27.3 in December. Siri, Writing Tools and Visual Intelligence will come under the control of a Chinese engine. Siri AI, the rebuilt Siri presented at WWDC 2026, remains off the radar for now.

This is a geopolitical moment as much as a product one. Apple has just set a precedent that the entire Western industry will be studying: to sell AI in China, you need a Chinese AI. What remains to be seen is what users gain — and what they agree to give up in exchange.

The Essentials

• Developer beta on November 4, public rollout with iOS 27.3 in December (source: press release from Apple China).
• Qwen powers Siri, Writing Tools, and Visual Intelligence in mainland China — an absolute first for Apple, which had never before entrusted its entire consumer AI to a third-party model.
• Siri AI is excluded from this first wave: China will receive the "Apple Intelligence 2024" experience, not the agentic Siri.
• Alibaba is contractually committed to not training Qwen on Apple users' requests, according to the terms of the agreement made public on October 31.
• The deeper motive is regulatory: mandatory CAC approval and data residency in China automatically disqualified any Western model.

Recommended Tools

Want to try out the engine that will power hundreds of millions of Chinese iPhones, or compare it with local competition? Three links are all you need:

What Apple actually announced

Two phases: a developer beta on November 4, then a public rollout in December with iOS 27.3. This is Apple's usual sequencing, but at an unusual pace — barely two months separate the announcement from the stable release, a sign that this project had long been maturing behind the scenes.

The announcement from Apple China lists three feature families:

• Siri: improved language understanding, plus writing and rephrasing assisted by Qwen.
• Writing Tools: rewriting, correction, summaries — Apple Intelligence's most mature module, and the easiest one to plug into a third-party LLM.
• Visual Intelligence: visual analysis through the camera, with object recognition, instant translation, and contextual search.

A telling detail: Apple talks about "Apple Intelligence features," never about a "new Siri." The wording is an admission. Chinese users will receive the 2024-2025 Apple Intelligence experience, powered by a 2026 engine.

Our fellow publication The Diary of AI confirms this timeline and stresses the same point: the absence of Siri AI in the first wave is not a communications mishap, but a regulatory red line.

On the infrastructure side, requests are routed through Alibaba's cloud infrastructure, hosted in China, in compliance with local data residency rules. Private Cloud Compute — the verifiable privacy bubble Apple uses in the West — has no announced equivalent here.

Finally, compatibility: the hardware matrix remains that of Apple Intelligence, meaning iPhone 15 Pro and later, as well as iPads and Macs running on Apple Silicon chips. No surprises on that front.

West vs China: two Apple Intelligences with nothing in common

No, it's not the same service under two flags. Model, infrastructure, feature catalog, and privacy regime differ from end to end. Here's the snapshot:

This table deserves a careful read, because it sketches out two philosophies. In the West, Apple pushes openness: the user chooses their brain. In China, the State chooses, and Apple executes.

The least-commented fault line concerns the Extensions. Siri's multi-provider hub makes no sense in a market where the list of approved models is set by the CAC. Offering Claude, Gemini, or Grok as a choice for a Chinese user isn't a feature — it's a legal impossibility.

Why Qwen? One word: regulation

Because no Western model can legally power a consumer AI service in mainland China. Since August 2023, China's generative AI regulations require certification from the CAC (the cyberspace administration), strict content alignment, and data residency within the country. The models from the American giants simply aren't offered there.

Apple had explored other avenues before landing on Alibaba. As early as 2024, the American press reported discussions with Baidu. In February 2025, Joe Tsai, Alibaba's chairman, publicly confirmed that his company was collaborating with Apple on AI in China. The matter thus matured for nearly two years — a timeframe that, in retrospect, looks like a negotiation as much political as it was technical.

That leaves the awkward question: why Alibaba rather than DeepSeek, Moonshot, or Z.AI, whose models compete at the very top? Because Alibaba checks three boxes that its rivals don't all check at the same time: a sovereign cloud infrastructure at national scale, a family of models covering the full spectrum from embedded to datacenter, and a business relationship documented since 2025.

My analysis doesn't beat around the bush: this isn't a benchmark decision, it's a market access decision. Apple traded the purity of its AI stack for a market where its Greater China segment accounted for roughly $67 billion in revenue in fiscal year 2024 alone (Apple annual report). Faced with Huawei's resurgence and state subsidies for local smartphones, the trade-off was mathematical.

Siri without Siri AI: China gets the packaging, not the new engine

Siri AI, the rebuilt Siri unveiled at WWDC 2026, is not part of the Chinese launch. Users in mainland China will get a Siri enhanced by Qwen, but not the agentic version — personal context, on-screen awareness, action execution — that is arriving in the West.

The reason is as much architectural as it is political. In the West, the rebuilt Siri runs on Gemini, as we covered when Ternus took the helm at Apple. But Gemini cannot operate in China. Porting the new Siri to Qwen would require revalidating the entire agentic chain with the CAC — months of procedures for an uncertain outcome. No date has been announced, and Apple has not promised that one ever will be.

The comparison with the West is stinging. There, iOS 27 transforms Siri into a multi-vendor hub where everyone chooses between Claude, Gemini or Grok (our analysis). In China: a single vendor, drawn from a state-approved list. Two visions of the personal assistant are clashing — consumer choice on one side, a single regulated offering on the other.

Should we feel sorry for them? An important nuance: Chinese users are going from zero system AI to a functional system AI. Writing Tools and Visual Intelligence will work properly, and written Chinese has historically been Qwen's strongest language — Alibaba built its reputation on it. The deficit is real, but it concerns the agentic capabilities, not basic assistance.

Qwen versus other Chinese models: a reasoned choice, not the best score

Objectively, Qwen is not the best-performing Chinese model at the moment. In our ongoing benchmark of open source LLMs, DeepSeek V4 Pro (Max) leads with a score of 88, ahead of Kimi K2.6 (85) and GLM-5.1 (83). The best Qwen3.6 variants top out at 74.

So why Qwen? Because Apple didn't buy a score: it bought a complete lineup. The Qwen3.6 family covers the entire spectrum, from the on-device model to the datacenter powerhouse — we detailed its architecture at launch.

The architecture, moreover, tells the strategy better than any press release. Apple didn't specify which variant runs where, but the signs are clear: a MoE with 3 billion active parameters like the Qwen3.6-35B-A3B is tailored for on-device use — controlled latency and power consumption on Apple Silicon — while the larger variants can handle complex queries on Alibaba Cloud. This is exactly the on-device + cloud scheme that Apple applies in the West with its own models.

One last argument, rarely highlighted: Qwen is the most widely deployed open source family in the Chinese ecosystem, with more than 100,000 derivative models claimed as of April 2025 (Alibaba). At the scale of hundreds of millions of devices, the maturity of an ecosystem carries more weight than a benchmark point. Since the DeepSeek quake of January 2025, the market has understood this well: distribution beats raw intelligence.

Privacy: Alibaba's Commitment, Taken at Face Value

Alibaba is contractually committed to not training Qwen on Apple users' requests. This is real, but it is narrow. The information, confirmed as early as the October 31 announcement, defuses the number one objection to such a partnership.

What the commitment covers: training. Your requests will not be used to improve future Qwen models.

What it does not cover: log retention periods, regulatory access, and the Chinese legal framework itself — the national security laws (2017) and data security laws (2021) impose cooperation obligations on Chinese companies that no commercial contract can neutralize. Equating this arrangement with the Western Private Cloud Compute, with its guarantees that are verifiable and audited by independent researchers, would be dishonest.

Yet it is important to keep a sense of proportion. For a Chinese user, the alternative was never "Apple Intelligence with PCC guarantees" — that option never existed there. The real alternative was the absence of Apple Intelligence, or local assistants that, for their part, train extensively on user data. Within that frame of reference, Alibaba's commitment represents a step forward, not a step back.

Apple's CEO, moreover, embraced this pragmatism in remarks reported by TechCrunch on October 31: no AI "for the sake of AI," but real features adapted to each market. A discourse consistent with the line set since Ternus took over at Apple — fewer spectacular promises, more actual deliveries. One may find this disappointing. It is above all realism.

What This Changes for Users and Developers

For Chinese users: a free system-level AI, finally. For developers: a new market, with a different engine and different rules.

On the consumer side, there's nothing to configure: Chinese Apple Intelligence arrives with iOS 27.3 on compatible devices, with no third-party account or subscription required. The catalog is shorter than in the West — no Siri AI, no multi-provider Extensions — but the essential building blocks are there, and they work in native Chinese.

On the developer side, it's more subtle. If you use Apple's Foundation Models framework, your calls will go through Qwen in China: a different model, a different generation style, a different tokenizer, different moderation. Three precautions are in order:

1. Test against the Qwen API before submitting any Chinese release. The outputs are not interchangeable with those of Western models, especially on structured formats.
2. Anticipate stricter moderation: the categories of acceptable content differ, and refusals will be more frequent on certain sensitive topics.
3. Rethink information retrieval: if your app needs to reason over documents rather than retrieve passages, the architectures are evolving fast — our feature on PageIndex, the vector-free RAG that reasons instead of searching explains why 2026 marks a turning point on this front.

And finally, a broader signal: Apple has just written the playbook that Meta, Microsoft and the rest will have to follow to enter China. A local partner, an approved model, a written commitment on data. The precedent has been set, and it will be cited in every negotiation to come.

❌ Common mistakes

Mistake 1: Believing "Apple Intelligence" refers to the same service everywhere

That ended in December. Different model (Qwen versus Apple and partner models), different features (no AI Siri in China), different infrastructure (Alibaba Cloud versus Private Cloud Compute). The solution: read the release notes by region and don't generalize any test performed elsewhere.

Mistake 2: Confusing "no training on my requests" with "guaranteed privacy"

Alibaba's commitment covers model training, nothing more and nothing less. Log retention, regulatory access, and the Chinese legal framework remain outside its scope. The solution: treat Chinese Apple Intelligence as a cloud service subject to Chinese law — because that's exactly what it is.

Mistake 3: Expecting AI Siri or Extensions in China in December

Both are excluded from the first wave, with no announced date. Porting agentic Siri to Qwen and getting it approved by the CAC will take months, maybe years. The solution: watch the iOS 27.3x betas, where these features would appear first if they were ever to arrive.

Mistake 4 (developers): Assuming an AI feature behaves identically on Qwen

A prompt calibrated on Western models will produce different outputs in China: style, length, formats, refusals. The solution: a dedicated test suite for the Chinese market, run against the Qwen API before every App Store update.

❓ Frequently Asked Questions

Which Qwen model exactly powers the Chinese Apple Intelligence?

Apple hasn't detailed the breakdown. Most likely, based on the architectures: an MoE variant with a reduced number of active parameters (like Qwen3.6-35B-A3B) on-device, and larger models on Alibaba Cloud for complex queries. Apple China's announcement mentions "Qwen" without specifying a version — the November 4 beta should shed more light on this.

Does Apple Intelligence cost anything in China?

No. As in the West, it's included for free on compatible devices, with no subscription or third-party account. Alibaba is paid by Apple on the server side, not by the end user: the business model remains that of a hardware selling point, not a subscription service.

Can you enable Qwen in Siri outside China?

No, the rollout is region-locked. Western users have their own stack: Apple's in-house models and a choice of providers via the iOS 27 Extensions (Claude, Gemini, Grok). Qwen isn't among them, and nothing suggests it ever will be — that's the political reciprocity of the deal.

Does Qwen remain open source despite this partnership?

Yes. The Qwen model weights remain published under the Apache 2.0 license, and the version integrated into iOS derives from this open family. But what Apple uses in China is a managed service on Alibaba Cloud, potentially adjusted and approved — not the raw open source weights you would download yourself.

Will Siri AI ever arrive in China?

No date yet. Two obstacles stack up: Gemini, its Western engine, cannot operate in China, and a Qwen variant of agentic Siri would need to be fully approved by the CAC. A "Chinese" Siri AI on Qwen is plausible in the medium term, but nothing has been announced.

✅ Conclusion

Apple has just proven it knows how to decouple its AI from its identity: same interface, Chinese engine under the hood, and a market of 1.4 billion people unlocked in exchange for a deliberate compromise. To learn everything about the model family that made this deal possible, head over to our deep dive on Qwen3.6.