苹果 AI 在中国测试千问，腾讯推出带视觉推理的 Hy4，中国 AI 芯片供应持续紧张

China AI Weekly Digest summarizes the week’s big news, model updates, compute, policy, products, and research, drawn from the China AI Index newsfeed. The digest is (mostly) compiled and drafted by an AI agent using credible news outlets; I fact-check and edit it. Read to the end for “One to Watch” as this week’s video pick.

This week: 181 stories tracked (2 editor picks, 59 English-language, 120 Chinese-language 🇨🇳) over 2026-10-30–2026-11-06.

The Big Three

Apple begins testing Qwen inside Apple Intelligence in China. Apple opened developer testing this week ahead of a planned December rollout with iOS 27.3. Alibaba’s Qwen will be available through Siri, Writing Tools and Visual Intelligence, giving users a way to ask about a photo or document without starting in a separate chatbot. The integration is opt-in, and users can choose what to send to Qwen. For Alibaba, the distribution matters as much as the model: Qwen gets a place in everyday iPhone workflows, even when users never open its app. (SCMP/雷峰网)

Tencent brings vision and computer use to Hy4, with WorkBuddy as its proving ground. The November 4 release adds visual reasoning and computer use to Tencent’s flagship model. In a hands-on test, Chinese media ifanr gave WorkBuddy a gameplay recording and used Hy4 to build a playable browser prototype. Users can show the agent what they want it to build. Tencent still acknowledges a gap with leading proprietary systems on some difficult research and coding tasks. Hosted access started with the launch; Tencent said the final weights and system cards would follow. (SCMP/爱范儿/@TencentHunyuan)

Chinese AI buyers run into separate bottlenecks for overseas GPU rentals and chip deliveries. The FT reports that Oracle dropped an estimated $7 billion, five-year deal to give Tencent access to as many as 100,000 AI chips in Southeast Asia after US officials raised concerns. Separately, CNBC reports that Moonshot explored renting Nvidia Blackwell capacity in Europe; those talks don’t establish a signed lease. Bloomberg reports that Chinese buyers expect about 1% of roughly $70 billion in RTX orders over the next four months. (FT/CNBC/Bloomberg)

Models

Baidu and Moonshot published weights, while Ant and Meituan introduced models built for longer tasks involving images and tools.

Baidu released ERNIE 5.5 Flash Base, a mixture-of-experts model with 280 billion total parameters and 14 billion active per token. Baidu says it supports up to a million tokens during fine-tuning and inference. This is a base checkpoint developers can adapt for their own tasks. (量子位)
Moonshot published the Kimi K3.1 Pro weights on Hugging Face this week, following the October launch through its own products and API. The download gives developers a self-hosting option for the 2.8 trillion-parameter model. (@Kimi_Moonshot)
Ant announced Ling-3.2-flash, a multimodal model with 560 billion total parameters and 25 billion active. Its stated targets include complex visual tasks, 3D interaction and medical reasoning. (@AntLingAGI)
Meituan launched LongCat-2.5, a natively multimodal model with 1.6 trillion total parameters and about 48 billion active. The company is targeting tasks that take many steps, including coding projects where the agent needs to inspect visual output as it works. (雷峰网/@Meituan_LongCat)

Compute

Domestic compute news spanned model training, server memory, optical links and chips for running models on devices.

DeepSeek plans to pretrain its next model on Huawei Ascend chips, Chinese media reports, citing supply-chain sources. Pretraining would put Ascend to work earlier in the model-development process than inference deployments do. (财联社)
CXMT plans to send customers samples of server DRAM built on its 4F² architecture by year-end. This is a separate development from its HBM3E effort. CXMT hasn’t set a mass-production date and says commercialization depends on customer validation and demand. (SCMP)
Hua Hong plans to expand silicon-photonics capacity as part of its next major expansion in 2027. Chairman Bai Peng singled out components for optical modules and power-management chips, two areas where the foundry wants to add capacity. (SCMP)
Houmo Intelligence reportedly closed a Series B worth several billion yuan. The chipmaker plans to expand deliveries of its M50 edge inference processor and develop its next compute-in-memory products. Houmo says the M50 has been selected for more than 150 customer designs, a measure of adoption that precedes shipments. (雷峰网)
Moffett AI closed a Series C+ of nearly RMB1 billion, according to Qbit. The chipmaker specializes in sparse inference and has adapted two remote-sensing models to its accelerator cards with a Peking University research partner—a specific application for hardware designed to skip unnecessary calculations. (量子位)

Funding

The robotics rounds backed specific industrial jobs: moving goods around sites and loading and unloading them.

Yicheng Innovation completed a RMB200 million Series C for its industrial autonomous-vehicle business. Its vehicles move materials around sites such as factories, logistics parks and ports. The company plans to use the proceeds to expand manufacturing capacity, improve the technology and deploy more vehicles. (投资界)
Qingtian Intelligence closed a Pre-A round worth tens of millions of yuan for robots that load and unload goods. Its industrial investors are also helping it test in operating logistics sites, giving the company a route from pilot installations to deployments across multiple facilities. (投资界)

Policy

Beijing signaled work on a broad AI law, while more targeted measures addressed employment and eldercare.

China’s National Development and Reform Commission said it would accelerate work on comprehensive AI legislation and develop safety standards and ethical rules covering the technology’s lifecycle. The November 3 briefing signals a legislative priority; China has yet to enact a unified national AI law. (财新)
Six departments issued measures encouraging the use of embodied AI, language models, brain–computer interfaces and exoskeletons in eldercare. The measures focus on developing uses for age-friendly products and services; they do not establish clinical effectiveness or authorize any particular device. (财联社)
Hello China Tech puts the State Council’s recent AI employment guidance alongside a Hangzhou dismissal case. The guidance encourages advance notice and help moving workers into new roles. As the analysis points out, protecting workers with existing contracts leaves a separate question: how many entry-level jobs companies will offer in the first place. (Hello China Tech)

Products

Agents showed up in phone demos, research desks and even a Singles’ Day livestream.

DeepSeek showcased Creator Mode in Harness, where users can describe a plugin and have the agent build it for use inside the app. The company shared prompts for seven demos, including a design workspace and a sandbox game. Creator Mode is distinct from the Super-agent research prototype below. (Bloomberg/投资界)
Doubao Work removed its five-hour usage cap for personal subscribers on November 6. Weekly allowances still apply. The change lets someone spend more of their existing allowance on a long task in one sitting without waiting for the shorter window to reset. (雷峰网)
StepFun’s device arm showed the STEPX Neo agent phone at Qualcomm’s booth at the China International Import Expo. Demos included checking Alipay offers, making travel videos with Jianying and planning trips through Ctrip. The expo appearance was a demonstration; retail availability wasn’t established. (量子位)
Alibaba Cloud introduced Qovest, a workbench for investment research. It retrieves data, gathers sources, generates reports and can revisit a subject on a schedule, keeping the results in a shared research workspace. Individual access began through invitation testing; an enterprise edition is planned. (雷峰网)
Livestream host Li Jiaqi sold Qwen Office token packages during the Singles’ Day campaign. Leiphone reports that the packages represented two trillion tokens sold in half an hour. That’s prepaid capacity sold through a consumer retail channel; it doesn’t measure how much customers have used. (SCMP/雷峰网)

Research

Teams explored ways to coordinate agents, rehearse rare road hazards and make robots useful across more tasks.

DeepSeek’s Super-agent prototype uses a router to assign work to agents that can act in parallel and respond to events across tools. In one internal example, a bug discussion triggered an attempt to reproduce the issue and a proposed GitHub fix for human review. The team has published parts of the design as a prototype. (SCMP/爱范儿)
Baidu’s Apollo World generates simulated drives so its robotaxis can practice rare hazards, such as a cyclist running a red light or an animal bolting into traffic. The system gives Baidu a way to stage encounters that are difficult to collect safely on real roads. (SCMP)
The Information reports that Meta and Shanghai-based Fourier built a humanoid prototype and are working on a second version. The collaboration gives Meta hardware on which to test its robotics ideas. The report describes internal development and identifies no commercial launch. (The Information)
The Chinese University of Hong Kong plans to begin a free dementia-screening program in Wong Tai Sin in January. It aims to reach 2,700 older residents over two years, combining AI-assisted retinal imaging with cognitive assessments to identify people who may benefit from follow-up. (SCMP)

One to Watch (NSFW warning)

Subscribe now