Apple Intelligence to Launch China Beta in Late October: Baidu and Alibaba Join as AI Race Extends to Devices and Robots

Apple Intelligence's entry into China is entering the beta countdown, with Baidu and Alibaba respectively entering search, Siri and model capabilities, while Apple points to December for the official rollout. Meanwhile, model services, short-drama traffic, edge hardware and humanoid robots are accelerating toward deployment in the same window, and AI competition is shifting from single-point model capabilities to system integration, user gateways and real-world execution.

Apple Intelligence in China: Beta First, Baidu and Alibaba Split the Gateway

Bloomberg reported that Apple plans to officially launch Apple Intelligence in mainland China through an iOS 27.3 update in December. The first batch will be AI features released in the US two years ago, covering iPhone, iPad and Mac, and will not include the new Siri AI assistant launched this year. The service had previously passed a generative AI service filing with cyberspace authorities on July 8 and obtained launch approval on September 29. Baidu is said to be involved in search and Siri-related feature cooperation; a module named Baidu Visual Search once appeared in the iOS 27 Beta 2 firmware. Alibaba, as an AI capability integrator, will provide Chinese users of iOS, iPadOS, macOS and visionOS with text and image understanding, content generation and other experiences, with the license covering Tongyi Wanxiang and Qwen. ^[1]

On October 15, Cook confirmed in Shanghai that Apple Intelligence and Alibaba's Qwen large model had been approved and would release a beta version to Chinese users later in October. The first batch includes real-time translation, writing tools, photo cleanup tools and more; Siri AI is not within this release. Bloomberg's Gurman previously reported that Apple plans to officially push it with iOS 27.3 in December, and the first batch is basically consistent with the version first released overseas in 2024. On the same day, Minister of Commerce Wang Wentao met Cook, and the two sides exchanged views on China-US economic and trade relations, AI development, Apple's business in China and other topics. ^[2]

Model Services: Outage Compensation and Developer Gateway Expansion

DeepSeek's web and API services experienced an outage on the morning of October 15. The API was unavailable from 8:11 to 13:04 and returned to normal after nearly five hours of repair. The official open platform issued a compensation notice: based on the account's consumption amount within seven days (October 8 to 14), the full amount will be returned in the form of bonus credits. The bonus credits arrive automatically, and API billing deductions are first taken from the bonus account. DeepSeek said it has launched a service stability review, sorting out and isolating single-point dependencies. This was its largest service problem this month, and related topics once trended on Weibo. ^[2]

TokenRhythm announced that more than 300,000 developers have joined, and launched a Double 11 presale limited-time benefit: users who make a single recharge of at least RMB 50 on the TokenRhythm API platform can receive an equal amount of model invocation credits, using one API Key to call all models on the platform; during the event, users can participate multiple times, with no cap on cumulative bonus credits, and bonus credits are permanently valid. The referral program launched during the same period offers invited new users RMB 18 in credits. The platform positions itself as a Chinese version of OpenRouter, displaying model configurations, prices and capability evaluations through a unified gateway, and pairs with its self-developed open-source agent OpenSquilla for multi-model intelligent fusion routing. Its API began public beta in August 2026 and signed strategic cooperation agreements with Infinigence AI and Alibaba Cloud in September. ^[4]

Capital and Growth: Anthropic Growth Cools, Tesla-SpaceX Merger Still a Forecast

Anthropic disclosed at an investor presentation for potential IPO investors that third-quarter revenue reached $16.8 billion, up about 45% quarter over quarter, and more than 3.6 times its total 2025 revenue; but compared with the more than 140% quarter-over-quarter surge in the second quarter driven by hot demand for Claude Code, growth slowed markedly this quarter. The company also disclosed that, as of the latest quarter, this year's adjusted operating margin was 11%. People familiar with the matter said 5% to 7% of Anthropic's revenue comes from the model distillation business, and this portion is expected to decline, continuing to drag on revenue in the current cycle; some investors attending the meeting believed the company needs to rein in valuation expectations. ^[2]

Barclays analyst Dan Levy expects that Musk may announce a merger of Tesla and SpaceX before December to integrate his AI, robotics and space businesses and strengthen control. The most likely plan is for SpaceX to acquire Tesla through a share swap; based on the companies' current valuations of about $2.1 trillion and $1.5 trillion, Musk could hold about 37% of the new company but control about 56% of voting rights. The transaction remains a Barclays scenario forecast; the two sides have not announced a formal merger plan. Potential synergies include providing Starlink communications for Tesla vehicles, integrating Grok into Optimus robots, and using energy storage products to support space data centers. ^[2]

Traffic Gateways: Tencent's Green Bubble Short Drama Takes Aim at Hongguo

Short dramas are becoming the internet's most lucrative time-spent gateway. QuestMobile data shows that in July 2026, Hongguo Short Drama's DAU reached 168 million, exceeding the combined DAU of Tencent Video, iQIYI, Youku and Mango TV, the four major long-video platforms; users spent an average of 125 minutes per day watching dramas, once exceeding Douyin main app's per-user usage in the same period. According to third-party monitoring data, by October its DAU had climbed to around 220 million, a net increase of more than 50 million in half a year. Hongguo has established a free ad-supported short drama model: Douyin provides the traffic gateway, Fanqie Novel provides IP reserves, and Hongguo handles long-duration viewing and completes ad monetization; in August this year, ByteDance products' total time share among the top 50 apps by monthly usage in China reached 41.8%, a record high. ^[3]

Tencent chose to use the WeChat ecosystem for a defensive counterattack. Green Bubble Short Drama, upgraded and renamed from the original WeChat official short drama mini-program, launched large-scale promotion in late September. Users do not need to download it and can use it by searching; it supports floating mini-window playback, Channels can directly mount and jump to it, and the homepage features Friends Watching and Regional Trending lists. On September 23, WeChat Open Platform and WeChat Marketing launched the Q4 2026 micro-drama mini-program incentive campaign. The core threshold is effective viewing time per user of no less than 10 minutes, and high-quality new dramas can receive up to 100% revenue incentives; during the National Day holiday, daily ad creatives reached a maximum of 7,101 sets. The gap is also clear: among the top 100 on Hongguo's overall list, short-drama popularity is generally above 40 million; the No. 1 drama on Green Bubble's overall list has popularity of about 21.19 million, and many dramas on the platform are old dramas already launched on other platforms. ^[3]

Edge Hardware: OPPO AI Xinli Ball Turns Recording into Long-Term Memory

OPPO AI Xinli Ball first appeared at the OPPO Developer Conference on September 17 this year. The main unit, including the rear clamping C-bridge, weighs 9.8 grams; placed in a battery-free carrying case, the whole thing weighs about 45 grams. Clamping force is above 4N, and it has IP55 dust and water resistance. It works all the time: in the morning, when taken out of the case and worn, it automatically enters listening and recording; at night, when put back, it automatically stops. A full charge supports 24 hours of continuous recording, with 5-meter long-range audio pickup and voiceprint recognition. Long-pressing the AI button on the front for 0.5 seconds wakes voice input; short-pressing marks the previous segment as a key point. The App's Today page automatically generates structured summaries and to-dos, and the PersonaX memory symbiosis engine also groups multiple conversations on the same topic into a project timeline. ^[5]

In response to privacy concerns about all-day wearing, OPPO keeps the front AI privacy light solid white while recording and, by design, keeps only text and not original audio; after transcription, only text records are retained, with full-link encryption through the OPPO AI PCC private cloud. On the software side, Xinli Ball is equipped with the O-Mind behavior insight algorithm and is positioned as an AI growth coach that can extract communication habits and improvement suggestions from long-term records. In external workflows, it supports persistent Q&A in WeChat and Feishu, and can also provide summary notes, inspiration notes and long-term memory through the MCP protocol for tools such as WorkBuddy, Codex, Claude and Cursor to call. The current version must be used with OPPO, OnePlus or Realme phones, mainly cooperating with the phone through the standalone Beluga AI App, and has not yet gone deep into the system notification layer. ^[5]

Embodied Intelligence: Humanoid Robots Move from Standing Firm to Whole-Body Work

The awkwardness of humanoid robots over the past few years is that running and jumping demos are lively, but to actually work, they first have to find flat ground, stand firm and then reach out. The underlying reason is that bipedal control has long faced a dilemma between stability and precision: to be stable, the hands have to shake along; to make the hands precise, the whole body easily falls. Mainstream industry solutions are mostly built on Nvidia's open-source whole-body control benchmark SONIC, but in Yuance Future's comparative tests, when the robot reaches toward a low or distant place through whole-body coordinated movements such as bending forward and extending, the open-source baseline's hand end-effector tracking error is on the order of 10 centimeters; at the same time, SONIC's control logic is almost entirely tied to flat ground, so locomotion and manipulation can only be switched in patches. ^[6]

Yuance Future takes whole-body intelligence WBI as its core route, self-collects human whole-body interaction data, and develops a humanoid foundation model system adaptable to different bodies. Its self-developed S0 motion control foundation model reduces hand end-effector tracking error to the 2-centimeter range and expands the control dimension to non-flat ground with height differences. When climbing stairs, it dynamically coordinates lower-limb foot placement, torso posture and center-of-gravity transfer, and also improves intent tracking for large movements such as bending over. The high-precision real-machine demonstrations and DAgger correction experience accumulated by S0 further feed the S1 brain, which has long-horizon task understanding. In the demo, the robot can lift a fallen bicycle, press a rice cooker's start button, and move a step stool to climb and reach objects; the rice-cooking task strings together crossing floor height differences, taking rice and adding water, removing and placing the inner pot, closing the lid and pressing the button into one long-horizon operation. This route has been validated on two configurations, Unitree G1 and Booster Robotics T2, and the two robots also divide work and cooperate in a refrigerator organizing task. ^[6]

Shared Trend: AI Deployment Enters a Triple Race of Systems, Scenarios and Bodies

From Apple Intelligence's phased entry into China, to model service providers competing for developers with stability and price, to Tencent defending the short-drama gateway with WeChat's social chain, OPPO turning recording hardware into a long-term memory gateway, and Yuance Future enabling humanoid robots to do housework with their whole bodies, the focus of AI competition has shifted from single-model demonstrations to sustainable services, reusable gateways and verifiable execution. What is worth watching next is the functional boundary between Apple's beta and official rollout, the approval progress of Siri AI, and the stability of multi-model routing, social distribution and whole-body control among real users and real working conditions. ^{[1][2][3][5][6]}