Eight contenders vying for one final spot, seven different losing outcomes: What Apple's China-market AI knockout round has screened out is not the participant with the strongest technical capability.

Apple spent three years selecting an AI brain for China-market iPhones, interviewing eight candidates. On July 8, 2026, "Apple Intelligence" completed its filing with the Cyberspace Administration of China, and in August, Reuters confirmed that Apple had co-trained a China-exclusive large model with Alibaba. The reasons for eliminating seven candidates are far more informative than the reason for picking the final one: what has never held Apple back is the ceiling of model parameters, but compliance, interest boundaries, and the delivery capacity of 100-million-level concurrent traffic. The competition of large models has evolved from "who has a higher benchmark score" to "who is least like a competitor" — this is the real ticket to the second half of the AI game.

"Apple is extremely picky. It has reached out to many Chinese companies, and finally chose us." In February 2025, Joe Tsai said this candid line when confirming the cooperation at the World Government Summit in Dubai.

How picky is Apple? According to reports from multiple media outlets including Caijing and CLS, Apple has contacted eight manufacturers in China over three years: Baidu, Alibaba, ByteDance, Tencent, DeepSeek, Baichuan, Moonshot AI, and Zhipu AI. This candidate list almost condenses the core competitive landscape of China's current large model industry: Baidu, which was the most highly anticipated at the earliest stage, dropped out midway; ByteDance, which holds the largest C-end traffic, has inherent barriers to commercial cooperation; DeepSeek, which delivered stunning benchmark results, failed to enter the final review; Tsinghua-affiliated Zhipu AI, Tsinghua-affiliated Moonshot AI, and Wang Xiaochuan's Baichuan all stopped outside the final selection.

One winner out of eight, seven different ways to lose. But every elimination reason points to the same standard — and this standard is different from what most people assume.

01

Two-Year Gap and a Compliance Wall: Why Apple Has to Choose, and Why It Is So Hard to Choose

Let's go back to June 2024. At that year's WWDC, Apple released Apple Intelligence. Overseas users got the basic features with iOS 18 in October the same year, while users of China-market iPhones waited for two full years.

The two-year blank period came at a real cost. In 2025, Huawei returned to the top of China's smartphone market with a 16.4% share, while Apple was pushed to second place with 16.2%. The key to winning the high-end phone market never lies only in camera systems and chips — when all manufacturers take AI as their core selling point, the Huawei Mate 70 series supports call summarization and AI object removal from photos, the Xiaomi 15 series runs MiMo locally on the device, vivo's Blue Heart AI is embedded in the photo album, while "Apple Intelligence" in the settings page of China-market iPhones is still gray and unclickable. This gap hurts more than any parameter comparison.

So Apple had to make a choice, and quickly. But the threshold for selecting a Chinese large model company is not technology, but compliance. Domestic generative AI services must complete filing before going online. Apple's self-developed model runs on overseas PCC (Private Cloud Compute), which cannot pass the data outbound compliance check. In June 2025, Apple submitted its filing plan, adopting a dual-architecture of "self-developed on-device model + local cloud large model". This process lasted for 13 months — the procedure was completed on July 8, 2026, and "Cyberspace China" made the announcement on July 15, making Apple the first foreign enterprise approved to provide self-developed AI models in China. Looking back at the leaked and then deleted "Apple Intelligence China Version Specification Document" in the spring of 2026, it was just a curtain blown open after everything was ready.

What has held Apple back is never technology, but compliance. The first step of compliance is to select a local partner that regulators and Apple are both assured of — this partner cannot be a search company at the same time, cannot be a short-video giant, cannot regard the WeChat ecosystem as its lifeblood, and cannot be a startup team that cannot even support 100-million-level concurrent traffic.

02

Seven Candidates Eliminated, Seven Mirrors: Every Failure Comes From Identity, Not Benchmark Scores

Who exactly did Apple contact? Cross-verified by multiple sources, Apple interviewed eight Chinese AI companies over three years. The reason for each elimination is a mirror, reflecting not the strength of the model, but the several weights in the "Apple Standard" that are more important than benchmark scores.

Baidu: Got up early, but arrived late.

Baidu was the most popular candidate in 2024, and the two sides once had in-depth cooperation. Apple even paid out of its own pocket to subsidize Baidu in model training and iPhone adaptation. But on the technical side, Ernie Bot was exposed for failing to meet Apple's standards in personalized response and privacy data processing; the more fatal problem is the interest structure — letting Baidu answer Siri's questions is equivalent to handing over the search entry on iPhones to a search company. The entry is Apple's lifeblood, and this conflict has no solution. Finally, it is reported that Baidu only undertakes peripheral modules such as visual search and content review in the China version of Apple Intelligence, while the core generation capability is assigned to others.

Caijing revealed that Baidu's advertising revenue in Q1 2026 fell by 22% year-on-year, and new AI businesses accounted for 52% of its total revenue. It is at a transformation stage that needs endorsement from benchmark clients, but the former top AI seed did not even make the starting lineup.

ByteDance: Natural Barrier as a Commercial Competitor.

Doubao has 382 million monthly active C-end users in China (QuestMobile, first half of 2026), with strong technology and sufficient capital. But there is a structural problem between ByteDance and Apple: the in-app purchase revenue share. ByteDance and Apple have long had structural disagreements on in-app purchase revenue sharing.

In August 2024, 21st Century Business Herald revealed that some paid services of Douyin Mini Games once redirected to external payment channels to avoid Apple's 30% commission. Apple then pressured Tencent and ByteDance to block this loophole, and the game between the two sides continues to this day. In addition, ByteDance has always been conservative about in-depth cooperation with Apple — it not only has the historical burden of the "Apple Tax", but also has geopolitical concerns about deep binding with US capital. The superposition of structural commercial conflicts and low willingness makes cooperation difficult to advance from the very beginning.

DeepSeek: Technically Leading But Production Capacity Limited.

DeepSeek V3/R1 has caused a stir in the overseas AI circle, and its capability is widely recognized in the industry. But according to Zhong Wenzhe and reports from Caijing, the core reason why Apple gave up DeepSeek is that the team lacks sufficient manpower and service experience to support stable calls from hundreds of millions of devices.

The downtime incident during the peak traffic period of the Spring Festival further confirmed to Apple that what it needs is not only model performance, but also full engineering capabilities to stably serve hundreds of millions of devices, cope with the concurrent traffic level of Double 11, and pass the multi-level protection and filing compliance review. Under the conditions at that time, DeepSeek's capacity building and service system did not meet Apple's requirements, which may be one of the reasons why Apple finally turned to other manufacturers.

Tencent: WeChat is an Independent Kingdom.

The best scenario of Hunyuan AI lies in the WeChat ecosystem, but the relationship between WeChat and iOS has always been delicate — revenue sharing, external links, mini programs, cloud games, every issue is a historical burden. Letting Hunyuan be deeply embedded in Siri and system-level writing tools is equivalent to binding two strong ecosystems together, and Apple is fully aware of the risks. Coupled with the unresolved "Apple Tax" dispute, Tencent did not enter the final review in the end.

Baichuan, Moonshot AI, Zhipu AI also failed to reach the final stage of contact. The reasons inferred from comprehensive information are mostly similar: Baichuan is still growing in scale and service capacity; Moonshot AI's Kimi has strong long-text processing capability, but its engineering and compliance system has not reached the level supporting 100-million-level terminals; Zhipu AI's GLM series has strong technical strength and pure Tsinghua background, but its C-end distribution capability is weak, and its commercial identity has no complementarity with Apple. The three lost not unjustly, and their loss is very typical.

Understanding Apple's partner selection logic means you understand the real ranking of China's AI industry: Model benchmark scores are the facade, while service capacity, interest boundaries, and compliance qualifications are the core essence.

03

Alibaba is Not the Strongest in Qwen, But the "Least Troublesome": Wins in the Underlying Logic, and Wins in Clean Identity

Then why did Alibaba win? Joe Tsai's candid line only stated the result without explaining the reason: Apple has been picky for three years, why did it finally stop being picky? Breaking it down, Alibaba won because of three superimposed factors that are all valid at the same time.

First, a clean identity. Alibaba does not make mobile phones, does not operate short-video platforms, and does not rely on search advertising for revenue. It has no direct business conflicts with Apple. Choosing Baidu means competition over entry, choosing ByteDance means disputes over revenue sharing, choosing Tencent means conflicts over ecosystems, choosing DeepSeek means concerns about production capacity. Only Alibaba — e-commerce, cloud computing, payment, logistics — all its businesses are outside the boundary of Apple's hardware business. There is no bad blood between it and Apple, only complementary cooperation.

Second, controllable technology. Qwen follows an open-source path, and the Qwen series has long been one of the open-source model families with the highest download volume on Hugging Face worldwide. Apple can conduct customized deployment and private debugging, and control the model's service mode on its own. For a company like Apple that is obsessed with user experience and never gives up on the PCC security architecture, "controllable" is more important than "the strongest". This is also the victory of the open-source ecosystem: the standardized interface and community tool chain accumulated by Qwen make it Apple's "least unfamiliar" partner.

Third, ready-made compliance system. Alibaba Cloud's generative AI service completed the filing in the early stage, with mature systems for data localization, multi-level security protection, and content review. Apple does not need to go through the long approval process together. After the 13-month filing process was completed and Apple became the first approved foreign enterprise, this combined model of "local large model + data localization + joint training" will most likely become the general template for foreign technology companies to enter the Chinese market.

The form of cooperation is deeper than the outside world imagines. On August 14, 2026, Reuters confirmed that Apple did not simply access Qwen, but co-trained a China-exclusive large model with Alibaba — Apple's self-developed on-device model handles real-time response and privacy-sensitive tasks, while Qwen handles long text, multimodal, and complex generation on the cloud, running in a dual-track mode. This is not a supplier relationship, but a co-construction relationship; it is not just "changing Siri's brain", but rewriting the intelligent base of the China version of iOS.

Looking at a higher perspective: Qwen has 167 million monthly active C-end users, less than half of Doubao's 382 million. On the App level, Alibaba actually lags behind. But after being embedded into the system underlying, Qwen can reach more than 200 million high-value users of China-market Apple devices at zero cost — this is an entry that no amount of promotion budget can buy. The first half of the game is about who has higher App monthly active users, the second half is about who can be embedded into the system underlying. Apps will become outdated, but system-level entries will not.

But don't rush to write an award speech for Alibaba. In this cooperation, Apple is also a party eager to solve its own problems: the Chinese market was overtaken by Huawei, and the AI feature for China-market iPhones was blank for two years. It needs a successful landing even more than Alibaba. For Alibaba, the system-level entry is always decided by Apple — Qwen is selected today, but if its experience lags behind, there are problems with filing renewal, or a cleaner candidate appears, it may also be replaced. Apple's "picky" attitude over the past three years will not expire on the signing day. A VIP card is not a free pass to stay forever. Getting a seat at the table and keeping that seat are two different things.

For the entire industry, technical strength determines whether you can get a seat at the table, the cleanliness of your identity determines whether you can sit in the main seat, and your commercial boundary determines how long you can stay. The competition of large model companies has evolved from the model layer to the commercial identity layer. Once Apple's table is set up, it will not only seat one company — Baidu has kept a backup position for visual search. Whether the next batch of models can meet its picky requirements is the real sequel of this knockout round.

The capability is expected to be rolled out in batches with the official version of iOS 27 and the iPhone 18 series in September 2026. At that time, users will only perceive a smarter Siri, but cannot perceive the three years, eight candidates, and a knockout round behind it. And who will be the next to take a seat at the new table?

References:

Baidu's Q1 2026 financial report: advertising revenue down 22% year-on-year (Securities Times)

Joe Tsai confirmed the cooperation at the Dubai Summit (February 13, 2025, Sina / Caijing Magazine)

Zhong Wenzhe / NetEase reported that DeepSeek was eliminated due to insufficient service experience (February 2025)

21st Century Business Herald, August 2024: ByteDance's Douyin Mini Games bypassed Apple's in-app purchase revenue sharing

QuestMobile first half of 2026: Doubao has 382 million monthly active users, Qwen has 167 million monthly active users

Reuters, August 14, 2026: Apple and Alibaba co-trained a China-exclusive model for China-market iPhones, adopting dual-track architecture

Caijing, February 2025: List of eight contacted manufacturers and the logic for Baidu / Tencent / ByteDance / DeepSeek's elimination

Cyberspace China announcement on July 15, 2026, Apple Intelligence filing completed on July 8, 2026 (China.com / Sina)

This article is from WeChat official account "New Entropy" (ID: baoliaohui