阿里巴巴 Qwen4-Max 多模态模型首度亮相

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: 248 stories tracked (0 editor picks, 99 English-language, 149 Chinese-language 🇨🇳) over 2026-11-07–2026-11-14.

The Big Three

Alibaba discloses a Hong Kong listing application for its chip unit. Alibaba said on November 10 that T-Head had confidentially filed for a Hong Kong listing on September 11. T-Head designs chips for data-center workloads, making it part of Alibaba’s own effort to build out AI infrastructure. A listing would give investors a separate way to value that chip business while keeping it within the group, as Alibaba expects T-Head to remain a subsidiary. Final offering terms, including how much the unit might raise, are still pending. (SCMP/雷峰网/HKEX)

MiniMax and ByteDance give image generation a production workflow. MiniMax released H3 Image with open weights and made it available through its API, MiniMax Code and Hailuo. The model generates images at native 4K and accepts up to 16 references, giving users more control over which elements carry into the result. ByteDance began rolling Seedream 5.5 into Doubao, Doubao Work and Jimeng, with localized edits and support for generating images with transparent backgrounds. (量子位/雷峰网/@MiniMax_AI/@BytePlusGlobal)

Jensen Huang resigns as Nvidia faces renewed scrutiny over chip exports and mounting U.S.–China tensions. Nvidia said Huang stepped down as CEO and from its board effective November 11, with CFO Colette Kress taking over as interim CEO. The board has launched an independent investigation into allegations surrounding Huang and Nvidia’s China business. Huang’s departure could reshape the global AI landscape, particularly U.S.–China technology competition. A vocal opponent of restricting chip sales to China, Huang had argued that keeping Chinese developers within Nvidia’s ecosystem served American interests. His exit comes as bipartisan lawmakers push for tighter export controls, including potential restrictions on overseas cloud access to advanced AI chips. (Nvidia 8-K/Reuters)

Models

This week brought downloadable weights from Meituan, a new speech model from Tencent and previews from Baidu and Qwen.

Meituan published the LongCat-2.5 weights under the MIT license, giving developers a way to run or adapt the multimodal model themselves. Meituan lists 1.6 trillion total parameters and about 48 billion active per token, with the model built for tasks that combine visual understanding and agentic coding. (VentureBeat)
Baidu introduced ERNIE 5.5 Flash Preview, a separate release from ERNIE 5.5 Base. Its demonstrations focused on practical tasks, including generating interactive 3D scenes through code and handling knowledge-work assignments. (雷峰网)
Tencent released Hy Realtime, a Hunyuan speech model for live voice interaction. The model takes speech input and generates spoken responses, giving developers a model designed for conversational audio. (第一财经/IT之家)
Alibaba previewed Qwen4-Max, the first model in its next-generation Qwen4 family, on November 11. The multimodal model supports text, image and video inputs, a 1 million-token context window, and reasoning capabilities designed for coding, complex agentic tasks and knowledge work. (@Alibaba_Qwen)
BytePlus announced Seedance 2.5 Flash and quoted US$0.152 per second for 720p text-to-video and image-to-video. It says Flash is nearly twice as fast as the flagship. (@BytePlusGlobal)

Compute

Server makers moved toward public markets while policymakers looked at who can buy advanced chips and where those chips end up.

Chinese AI server maker xFusion cleared a Shenzhen Stock Exchange review for its proposed ChiNext listing, Reuters reported. The planned offering is sized at about US$1.2 billion, and the review clears one stage of the listing process. (Reuters)
SMIC expects tight AI computing supply to persist until 2028, according to SCMP. The foundry’s forecast puts a timeframe on the capacity pressure facing AI developers as Washington weighs tighter controls on advanced chips. (SCMP)
Malaysia is considering closer oversight of data centers to curb the resale and diversion of advanced AI chips, Caixin reported. The proposal would bring data-center operators into the enforcement picture by focusing on what happens to chips after they enter the country. (财新)
U.S. Senator Tom Cotton urged Commerce Secretary Howard Lutnick to block Chinese DRAM maker CXMT from the U.S. market. The request targets access to U.S. buyers and would require action by the administration. (Reuters/CNBC)

Funding

Kimi’s maker took preparatory steps toward a listing, while two robotics rounds funded manufacturing and factory deployments.

Moonshot AI is restructuring ahead of a potential IPO, the WSJ reported. The Kimi developer is moving shareholder interests from a Cayman Islands structure into a Chinese entity. The Journal put its latest private valuation at US$50 billion. (WSJ)
UniX AI closed a RMB500 million pre-A round, with Greenland Group joining as a strategic investor. The Suzhou robotics company says the money will help it scale production and deliveries of its general-purpose humanoids. (投资界/新京报贝壳财经)
Zhifeng Chuangyuan raised RMB80 million in an angel round. The company has introduced its wheeled, dual-arm Z100 and plans to spend the money on R&D and expansion into optical-module manufacturing, where robots need to perform precise insertion and removal tasks. (投资界)

Policy

Beijing set a robot manufacturing target, Shanghai laid out rules for scientific data, and U.S.–China safety talks surfaced a basic question about incident reporting.

Beijing’s 2026–2028 embodied robotics plan calls for local production capacity of 200,000 finished robots a year. The city will support both contract manufacturing and companies’ own production lines, alongside shared facilities across Beijing, Tianjin and Hebei to speed up parts prototyping. (财联社/北京市政府)
Shanghai issued trial rules on preserving, classifying and sharing scientific data. The measures also encourage research institutions and AI developers to work together, putting scientific datasets within reach of more model-development projects where sharing is permitted. (第一财经/上海市科委)
Washington and Beijing established a channel to discuss risks from advanced AI. Analysts told SCMP that agreeing on what counts as a “serious” incident could be an early hurdle. That definition would shape which events either government is expected to bring to the other’s attention. (SCMP)

Products

DeepSeek opened up tools from its consumer app, while Qwen added an agent that colleagues can use together.

DeepSeek opened APIs for capabilities already used in its app, including search, document and link ingestion, voice and safety checks. Developers can also call standalone services for optical character recognition, speech recognition and speech synthesis. (第一财经/@deepseek_ai)
Qwen Office 2.0 introduced Team Agent, a shared workspace where colleagues can work with the same assistant. A group can bring its own materials into a shared task and collaborate on the output in that workspace. (量子位/雷峰网)
Apple completed Chinese generative-AI service filings covering tablets and laptops. The filings extend the registered device categories for Apple Intelligence. The report gave no rollout date for either category. (第一财经)

Research

Robot teams tested how touch and movement transfer across hardware and environments, while a new drug-discovery model took on biomolecular benchmarks.

Ant-backed Robbyant’s LingBot-Touch learns a common representation across pressure-sensitive, magnetic and camera-based tactile sensors. The team says it used nearly 5,000 hours of training data and evaluated the approach on physical hardware, testing how much robots can reuse across different ways of sensing contact. (量子位/arXiv)
Apple researchers tested HALO, a humanoid control system trained in simulation, on a Unitree G1 carrying a Mac mini. The setup ran a Qwen vision-language model locally and used the robot to test movement around furniture and other obstacles in unfamiliar spaces. (爱范儿/Apple Research)
Geodesic Intelligence, founded by former ByteDance scientist Quanquan Gu, released NovaAtom-Lite to predict biomolecular structures and interactions. Gu told followers that it led several tasks on a third-party benchmark. The reported scores cover computational predictions used in drug discovery. (SCMP)

One to Watch

Subscribe now