EastVoice · AI · 2026-11-05 · Issue #119

Editor's Note: JD put a number on its robot ambition at Wuzhen: 3 million robots, 1 million driverless vehicles and 100,000 drones purchased over five years, on top of 100,000 real work tasks already recorded. Bloomberg, meanwhile, puts China's robot funding this year at $1.6 billion and reports that investors are starting to talk about a shakeout. Tencent shipped Hy4 and promised the weights.

Commentary: The order book and the funding line rarely move in the same direction. This week they moved apart.

---

I. Humanoid Robots

1. JD wants three million robots in five years, and it already has 100,000 tasks on tape

• Summary: Speaking at the Wuzhen robotics forum, JD's tech committee chair Cao Peng said the company has collected more than 100,000 real work tasks across 100-plus production and service scenarios. Its five-year purchase plan runs to 3 million robots, 1 million driverless vehicles and 100,000 drones.
• Source: ITHome · 2026-11-04
• Editor's Take: Data collection is the part nobody can buy later. JD is turning its own warehouses into a training set, and the hardware order is the smaller half of the story.

2. Meta built its humanoid with a Chinese partner, and its VP camped on the factory floor

• Summary: The Information reported on Oct 31 that Meta co-developed a humanoid prototype codenamed Max with Shanghai-based Fourier. A Meta vice president spent close to two months inside Chinese factories during the build, and the project went from kickoff to prototype faster than the team had planned.
• Source: Huxiu · 2026-11-04
• Editor's Take: Meta's advantage in this category is capital and software, not machine tools. The two months on the floor bought something a purchase order cannot.

3. Six robots, 24 hours, one convenience store: four kept the job

• Summary: AgiBot put six of its G2 humanoids through a 24-hour shift in a live-streamed convenience store on Nov 2. More than 2 million people watched. Four of the six finished the shift and were kept on; the other two did not make the cut.
• Source: ITHome · 2026-11-04
• Editor's Take: A public failure rate of one in three is more informative than any highlight reel. Retail work is mostly reaching, and reaching is where the current generation still loses.

4. China's robot funding is $1.6 billion this year, and the talk has turned to a shakeout

• Summary: Bloomberg puts 2026 funding for Chinese robotics companies at $1.6 billion and reports growing concern among investors that too many startups are chasing the same few deployment sites. The piece frames the slowdown as the start of consolidation rather than a loss of interest.
• Source: Bloomberg · 2026-11-04
• Editor's Take: Money is not leaving the sector, it is concentrating. The companies with signed deployments will raise again; the ones with demo videos will not.

5. A full-stack embodied model shop turns up at Hong Kong Fintech Week

• Summary: Kinetix AI (超维动力) brought its high-dynamic humanoid hardware and embodied foundation model to the Hong Kong Fintech Week x StartmeupHK 2026 at the HKCEC, pitching a full stack from model to body to deployment partners.
• Source: QbitAI · 2026-11-04
• Editor's Take: Fintech week is an odd venue for a robot. It tells you where this company expects the money to come from.

---

II. Autonomous Driving

1. RoboSense turns on a LiDAR line that runs with six people a shift

• Summary: RoboSense started what it calls the world's first fully automated digital LiDAR production line, at its Honghualing site in Shenzhen's Nanshan district. Each line needs six people per shift and is rated for 500,000 units a year. The building is also the country's first "industry upstairs" pilot.
• Source: ITHome · 2026-11-04
• Editor's Take: LiDAR economics have been stuck on assembly labor for years. Six people per line is the number that decides whether the price war continues.

2. Seres says it is getting ready for L3, without naming a car

• Summary: Seres told investors that L3 conditional autonomy depends on national standards, product approval, local scenario policy and compliance testing, and that it is preparing on all four fronts. It declined to name models or filing dates, saying those will come from official channels.
• Source: ITHome · 2026-11-04
• Editor's Take: Every automaker in China is "preparing for L3." The ones that matter will be the first to put a model number and a liability clause next to the claim.

3. Li Bin: the best-selling embodied agent on the road is a smart EV

• Summary: At Wuzhen, NIO's founder said the highest-volume embodied agent shipping today is the intelligent electric car, not the humanoid. He argued autonomy should free up a driver's attention rather than take work away from taxi and ride-hail drivers.
• Source: ITHome · 2026-11-04
• Editor's Take: He is right about the install base and careful about the politics. The second half of that sentence is aimed at regulators, not customers.

4. Dongfeng puts its own ADAS in a ¥159,800 SUV and drops Huawei

• Summary: Dongfeng's Yipai M8 Tianyuan Laser edition launched from ¥159,800, with the company's in-house driver-assistance stack replacing the Huawei Qiankun system used in the previous version. The car keeps the lidar hardware.
• Source: ifanr · 2026-11-04
• Editor's Take: A tier-one supplier losing a socket to the automaker's own team is the pattern to watch in 2027. Huawei's pricing power depends on how many of these switches happen.

---

III. World Models & Physical AI

1. MoWorld 2.0 runs on Ascend and lets you edit the scene it just made

• Summary: Moxin Technology released MoWorld 2.0, a 4D world model built natively on Huawei's Ascend stack, at the Wuzhen forum on AI and new productive forces. The company says generated scenes can be interacted with, edited and called by other systems rather than only watched.
• Source: ITHome · 2026-11-04
• Editor's Take: Editability is what separates a world model from a video generator. The Ascend dependency is also the point, given who the buyer is.

2. XPeng's 4B perception model beats GPT-6 Astra on 34 benchmarks

• Summary: XPeng, with Peking University and the University of Hong Kong, published GroundingPI, an embodied perception base model built on a "perception first" recipe. At roughly 4 billion parameters it ranked first across 34 public perception benchmarks and showed better data efficiency in robot simulation tasks.
• Source: ITHome · 2026-11-04
• Editor's Take: Four billion parameters beating a frontier model on perception is a useful reminder that these benchmarks measure a narrow skill. It is also exactly the narrow skill a robot needs.

3. A state-owned port group's lab ships a 35B embodied reasoning model

• Summary: The China Merchants Lion Hill Lab released Lion-ER-35B-A3B, a mixture-of-experts embodied reasoning model. The team argues that physical common sense does not come from adding data, and trains the model on cases where screen position and real 3D position disagree, or where the same object looks different from a new angle.
• Source: QbitAI · 2026-11-04
• Editor's Take: The insight is that a camera image is not a measurement. Most robot failures start with a model that quietly believes it is.

---

IV. Foundation Models

1. Tencent's Hy4 lands, and the weights are coming

• Summary: Tencent released Hy4, calling it the first Hunyuan model to rank among the strongest open models. The company says weights and a system card will follow shortly, with published work on network and biosecurity risk. The release lands in the middle of an open-weight race with DeepSeek, Z.ai and Alibaba.
• Source: ITHome · 2026-11-04
• Editor's Take: Announcing the weights before shipping them is now standard practice in China. The gap between the two dates is where the real capability gets revealed.

2. DeepSeek closes a ¥100 billion round with CATL and Tencent

• Summary: DeepSeek's second round is reported at roughly ¥100 billion (~$14 billion), with CATL and Tencent increasing their stakes; one account puts the final total at ¥105.5 billion. Either figure would be the largest single round raised by a Chinese AI company.
• Source: Huxiu · 2026-11-04
• Editor's Take: A battery maker is now one of the largest shareholders in a model lab. In a market where power is the binding constraint, that is not a strange pairing.

3. Huawei puts openPangu-2.1-Pro into Xiaoyi Work, then retires the 2.0 pair

• Summary: Huawei's assistant Xiaoyi Work is the first product to run openPangu-2.1-Pro, under a "HarmonyOS task enhancement" banner. Huawei Cloud separately notified customers that MaaS stops selling openPangu-2.0-Flash and 2.0-Pro on Nov 6, three weeks after the notice.
• Source: ITHome · 2026-11-05
• Editor's Take: Three weeks of deprecation notice is a hard cutoff for anyone who built on the 2.0 endpoints. Enterprise buyers now price that risk into every Huawei contract.

4. ByteDance tells staff to stop pasting company code into Claude and Codex

• Summary: An internal notice, reported on Nov 3, bans external AI access to all L4-level code repositories and Feishu documents company-wide starting this month. Employees who bought their own Claude or Codex accounts cannot use them on that material either.
• Source: ITHome · 2026-11-04
• Editor's Take: This is the first large Chinese company to treat coding agents as an exfiltration channel rather than a productivity tool. The internal alternative will be judged on whether it is good enough to stop the workarounds.

5. Alibaba sold 2 trillion tokens on a livestream for a penny a card

• Summary: On Nov 2, Li Jiaqi's livestream listed Qwen Office personal monthly cards at ¥0.01, down from ¥28. About 100,000 sold in half an hour, roughly 2 trillion tokens of quota. Alibaba vice president Chen Yusen appeared on the stream to sell it.
• Source: TMTPost · 2026-11-04
• Editor's Take: Token inventory is being merchandised like a beauty sample. It works as customer acquisition and tells you how much idle inference capacity Alibaba is sitting on.

6. Moxin raises nearly ¥1 billion for sparse inference chips

• Summary: Moxin AI in Shenzhen closed a C+ round of close to ¥1 billion, with Renai Capital, Haitang Fund, Dongheng Petroleum and others participating. The money goes into hardware and software for its sparse-inference architecture.
• Source: Huxiu · 2026-11-04
• Editor's Take: Sparse inference is a bet that most of a model's weights are dead weight at serving time. If it holds, the savings land directly on the token price.

7. Houmo quietly closes a multi-billion-yuan B round for edge AI chips

• Summary: Houmo Intelligence (后摩智能), which builds compute-in-memory chips for on-device models, closed a B round in the several billion yuan range, according to people familiar with the deal. The company has raised several times in the past twelve months.
• Source: Leiphone · 2026-11-04
• Editor's Take: Ten edge-AI chip rounds in September alone, by Leiphone's count. Capital is moving from the cluster to the handset, and compute-in-memory is the most credible way to get efficiency there.

---

V. AI Software

1. ByteDance stops selling Coze's enterprise tier while it rebuilds the product

• Summary: Coze published a notice that it is consolidating its enterprise products. The standard and flagship enterprise plans stop taking new customers from Nov 30, existing packages run until Dec 31, 2027, and unused balances are refunded after that. A separate notice from ITHome describes the move as a pause on new sales during the integration.
• Source: TMTPost · 2026-11-04
• Editor's Take: Killing a paid tier with a thirteen-month runway is an orderly retreat. The question is what replaces it, and whether enterprise agent builders trust the next contract.

2. Ant's Lingguang adds 40 connectors, from McDonald's to WeChat Read

• Summary: Lingguang 2.0 picked up podcast subscriptions, Quark Drive, Hello's ride-hailing, McDonald's and WeChat Read in its Nov 4 update. The personal agent now claims 40 external connectors, including Alipay, NetEase Mail and Feishu.
• Source: Leiphone · 2026-11-04
• Editor's Take: Connector count is the new app-store chart. It measures partnerships, not whether anyone books a ride through the agent twice.

3. Huawei Cloud's agent platform takes CAICT's top rating

• Summary: AgentArts passed the China Academy of Information and Communications Technology's trusted enterprise agent platform evaluation, becoming the first product to reach the top 4+ level.
• Source: ITHome · 2026-11-04
• Editor's Take: A government-backed rating is becoming a procurement requirement in state-owned enterprises. Being first matters more than the score.

4. China starts registering agent identities, and Ant is one of the first operators

• Summary: An agent identity registration and resolution platform launched at the Wuzhen summit, with Ant Group among the first batch of operating platforms. The framing from the conference is that governance is moving from content to behaviour, because agents now act on a user's behalf.
• Source: Leiphone · 2026-11-04
• Editor's Take: An identity layer for agents is the precondition for any liability regime. It also hands the registry operator a very good view of who is calling whom.

5. A film studio's AI production platform claims 30% off the cost of a minute

• Summary: Hujing Entertainment (虎鲸文娱) launched Jingrui AI, a production and management platform for film and TV. It bundles more than 50 third-party video models and 10-plus in-house ones, runs an infinite canvas rated for thousand-node scenes and hundred-person collaboration, and claims per-minute cost savings above 30%.
• Source: Leiphone · 2026-11-04
• Editor's Take: Studios are becoming model aggregators because no single video model is reliable across a whole episode. The margin sits in routing.

6. A Shanxi city gets an AI agent for parking spaces

• Summary: Shanxi unveiled Bolin (泊邻), which it calls the country's first urban parking AI agent, at the China Urban Parking Industry Conference. Built by Shanxi Construction Investment's static traffic unit, it merges live data from roadside bays, public lots and shared residential spaces into a heat map that planners can use to reallocate capacity.
• Source: ITHome · 2026-11-04
• Editor's Take: Municipal agents will be sold on measurable outcomes, and parking is one of the few where the outcome is a number. Expect more of these before any of them work well.

---

物界前沿 | EastVoice