Headlines · Hot List · Forum · 中文

EastVoice

PHYSIX FRONTIER EASTVOICE

2026.11.12 · Thursday

Issue 079

FIVE MINUTES A DAY, THE GLOBAL AI SHIFT

¥96.9 billion

Shangwei New Materials record market cap on humanoid trade

3 million

Ant Lingguang 2.0 hosted tasks in two weeks

¥1,000 per hour

Embodied-data middlemen quotes while workers earn ¥8-¥15

30B

Huawei Kirin 9050 on-device mixture-of-experts model

Top Stories 3 · dual commentary | Section Picks 25 · one-line takes

TOP STORIES

Shangwei New Materials hits a record ¥96.9 billion, Beijing offers embodied AI firms one-on-one IPO coaching, and Chengdu's fifth-generation Gongga humanoid claims the best brain in the business.

1A humanoid takes over the hotel laundry room

QbitAI profiles an embodied system now running laundry duty at a hotel: it collects dirty clothes from guest rooms, threads narrow corridors, and loads baskets. The stack pairs a slow deliberation model with a fast reflex model, plus reinforcement-style iteration on the machine itself.

What it means for you|The system now collects dirty clothes from guest rooms, threads narrow corridors, and loads baskets at a hotel, pairing a slow deliberation model with a fast reflex model.

Editor's take|Laundry is a better first job than most demos. The route is fixed, the payload is light, and the failure mode is a stained shirt rather than a crushed human.

Source: QbitAI (2026-11-11)

2Shangwei New Materials hits a record ¥96.9 billion on the humanoid trade

The stock touched ¥244.2 intraday on Nov 11 and closed at ¥240.31, up 9.8%, for a market cap of ¥96.9 billion. Third-quarter R&D spending rose 23-fold year on year. The company issued a risk warning two days earlier.

What it means for you|The stock touched ¥244.2 intraday on Nov 11 and closed at ¥240.31, up 9.8%, for a market cap of ¥96.9 billion, with third-quarter R&D spending up 23-fold year on year.

Editor's take|A materials supplier re-rating on robot narratives, with R&D up 23x and a risk notice already filed. That combination has a name in A-shares, and it is not "fundamentals".

Source: Huxiu (2026-11-11)

3Beijing offers embodied AI firms one-on-one IPO coaching

The municipal bureau of economy and information technology issued a 2026-2028 action plan for the embodied AI industry, including a dedicated service channel for companies heading to listing and support for building national champions.

What it means for you|The municipal bureau of economy and information technology issued a 2026-2028 action plan for the embodied AI industry, including a dedicated service channel for companies heading to listing.

Editor's take|Local governments have decided which part of the stack they can actually control. Not model quality — the queue to the exchange.

Source: ITHome (2026-11-11)

SECTION PICKS

25 AI stories worth your time, each with a one-line take.

Humanoid RobotsInside the embodied-data trade: ¥1,000 quotes, ¥8 an hour

An investigation into robot training data found middlemen quoting up to ¥1,000 per hour while the workers recording the footage take home ¥8 to ¥15 an hour. One contributor spends her shifts folding and shaking out jackets in front of a head-mounted camera.

What it means for you|Middlemen quote up to ¥1,000 per hour while the workers recording the footage take home ¥8 to ¥15 an hour, with one contributor folding and shaking out jackets in front of a head-mounted camera.

One-line take|Every robotics company's margin is somebody's piece rate. The data layer is being built on the cheapest labour China has, which is exactly why it looks so scalable on a slide.

Source: Huxiu (2026-11-11)

Humanoid RobotsChengdu's fifth-generation Gongga humanoid claims the best brain in the business

The Chengdu Humanoid Robot Innovation Center unveiled its fifth-generation Gongga, claiming the most advanced humanoid brain technology globally. Hardware details and shipping dates were thin.

What it means for you|The Chengdu Humanoid Robot Innovation Center unveiled its fifth-generation Gongga, claiming the most advanced humanoid brain technology globally, though hardware details and shipping dates were thin.

One-line take|"World-leading" without a benchmark number is a press release, not a product. Fifth generation in under three years tells you the iteration is on the slide deck, not on the line.

Source: ITHome (2026-11-11)

Autonomous DrivingFinland clears Tesla's driver-assist system, the 11th country in Europe

Finland's transport authority Traficom approved Tesla Assisted Driving for public roads, making it the 11th European country to green-light the system, after Latvia.

What it means for you|Finland's transport authority Traficom approved Tesla Assisted Driving for public roads, making it the 11th European country to green-light the system, after Latvia.

One-line take|Europe is approving Tesla one country at a time, and each approval is narrow. That is not a rollout, it is a paperwork grind with a map attached.

Source: ITHome (2026-11-11)

Autonomous DrivingMusk says the Cybercab will carry most passenger miles

Musk argued the Cybercab will handle the bulk of passenger travel in every jurisdiction that permits it, calling the conclusion obvious to anyone with eyes and a brain.

What it means for you|Musk argued the Cybercab will handle the bulk of passenger travel in every jurisdiction that permits it, calling the conclusion obvious to anyone with eyes and a brain.

One-line take|The claim is unfalsifiable until a permit exists. Tesla's bottleneck has not been the vehicle for three years; it is the regulator and the insurance market.

Source: ITHome (2026-11-11)

Autonomous DrivingWeRide turns homeward, pitching a Guangzhou-to-Hong Kong robotaxi

Founder Han Xu told CNBC at Hong Kong Fintech Week he wants passengers riding a robotaxi from Guangzhou to Hong Kong within four to five years — sleep in the car, wake up across the border.

What it means for you|Founder Han Xu told CNBC at Hong Kong Fintech Week he wants passengers riding a robotaxi from Guangzhou to Hong Kong within four to five years — sleep in the car, wake up across the border.

One-line take|WeRide built its story on overseas markets; now it is selling the Greater Bay Area to Chinese investors. Cross-border driving means two regulatory regimes, which is the hardest part of the pitch.

Source: Huxiu (2026-11-11)

Autonomous DrivingDongfeng Peugeot Citroën hands its cockpit to ByteDance's Volcano Engine

The joint venture signed a strategic deal with Volcano Engine covering AI transformation, co-developed product technology, and full-lifecycle customer service. New energy models will carry the latest AI cockpit stack.

What it means for you|The joint venture signed a strategic deal with Volcano Engine covering AI transformation, co-developed product technology, and full-lifecycle customer service, with new energy models carrying the latest AI cockpit stack.

One-line take|Legacy JVs have given up building their own assistant. Renting ByteDance's is faster, cheaper, and quietly turns the carmaker into a hardware vendor.

Source: ITHome (2026-11-11)

World Models & Physical AITed Xiao: verification, not generation, is what stalls AI in the physical world

Xiao, a founding technical member at Prometheus and formerly of Google DeepMind, argued at Humanoid Hub that closing the loop — reliably verifying whether an action worked — is the binding constraint on robots, not model capacity.

What it means for you|Xiao argued at Humanoid Hub that closing the loop — reliably verifying whether an action worked — is the binding constraint on robots, not model capacity.

One-line take|This is the least glamorous argument in robotics and probably the correct one. A model that cannot tell whether it succeeded will happily repeat the same mistake for a thousand episodes.

Source: Huxiu (2026-11-11)

World Models & Physical AILi Auto's OTA 8.7.0 adds a driving-style dial and glasses-controlled parking

Li Auto pushed OTA 8.7.0 to MEGA and i-series models, adding driving-style switching to its VLA stack and letting a paired pair of glasses pick the parking slot by pointing.

What it means for you|Li Auto pushed OTA 8.7.0 to MEGA and i-series models, adding driving-style switching to its VLA stack and letting a paired pair of glasses pick the parking slot by pointing.

One-line take|Shipping a VLA model as a user-facing toggle is the honest admission that end-to-end driving still behaves differently on different days.

Source: ITHome (2026-11-11)

World Models & Physical AITeaching robots where to grab, with 3D affordance in the data pipeline

A research group proposes injecting 3D affordance into manipulation data generation, so synthetic episodes encode where an object can actually be grasped instead of just where it appears.

What it means for you|A research group proposes injecting 3D affordance into manipulation data generation, so synthetic episodes encode where an object can actually be grasped instead of just where it appears.

One-line take|Most synthetic grasp data is geometrically valid and physically useless. Affordance is the cheap fix, and cheap fixes are what this field actually ships.

Source: AI Era (2026-11-11)

Foundation ModelsBaidu ships ERNIE 5.5 Flash Preview, weights open by Nov 25

Baidu released ERNIE 5.5 Flash Preview, the first post-trained version built on its Base model. The company says weights go open before Nov 25.

What it means for you|Baidu released ERNIE 5.5 Flash Preview, the first post-trained version built on its Base model, and says weights go open before Nov 25.

One-line take|A Chinese lab promising open weights before the model is even benchmarked publicly is a distribution play, not a research announcement.

Source: ITHome (2026-11-11)

Foundation ModelsAlibaba teases Qwen4, with the Max preview already live

Alibaba confirmed Qwen4 is coming, promising deeper knowledge, better spatial awareness, and stronger coding and agentic behaviour. Qwen4-Max is already available in preview.

What it means for you|Alibaba confirmed Qwen4 is coming, promising deeper knowledge, better spatial awareness, and stronger coding and agentic behaviour, with Qwen4-Max already available in preview.

One-line take|Qwen's real advantage was never the top-line score. It is that every size tier ships, and every size tier gets downloaded.

Source: ITHome (2026-11-11)

Foundation ModelsHuawei's Kirin 9050 runs a 30B MoE model on the phone

Huawei detailed the next-gen Da Vinci NPU in the Kirin 9050 series, built on a "logic folding" τ architecture, claiming it can run a 30B mixture-of-experts model on-device with no network and no cloud round-trip.

What it means for you|Huawei detailed the next-gen Da Vinci NPU in the Kirin 9050 series, claiming it can run a 30B mixture-of-experts model on-device with no network and no cloud round-trip.

One-line take|On-device 30B is a memory-bandwidth claim dressed as an AI claim. If it holds, it changes what a phone assistant can do offline, and that matters more than a benchmark.

Source: ITHome (2026-11-11)

Foundation ModelsAndroid's AI benchmark board adds GPT-6.1 Terra, Kimi K3.1 Pro, DeepSeek V4.1 Flash

Google's Android Developers account published the first scores for three models on its AI development benchmark, including Kimi K3.1 Pro and DeepSeek V4.1 Flash alongside OpenAI's GPT-6.1 Terra.

What it means for you|Google's Android Developers account published the first scores for three models on its AI development benchmark, including Kimi K3.1 Pro and DeepSeek V4.1 Flash alongside OpenAI's GPT-6.1 Terra.

One-line take|Google putting two Chinese models on its own developer leaderboard is the clearest sign that on-device AI is a procurement market, not a patriotism contest.

Source: ITHome (2026-11-11)

Foundation ModelsMoore Threads adapts MiniMax's H3-Image on day zero

Moore Threads said its MTT S5000 accelerator and MUSA software stack supported MiniMax's H3-Image image model on the day it was open-sourced.

What it means for you|Moore Threads said its MTT S5000 accelerator and MUSA software stack supported MiniMax's H3-Image image model on the day it was open-sourced.

One-line take|Day-zero support is the only number that matters for domestic accelerators. If a new open model needs a two-week port, developers go back to Nvidia and never return.

Source: ITHome (2026-11-11)

Foundation ModelsAlibaba open-sources Qwen Code 1.0, folding in its Qoder harness

Alibaba released Qwen Code 1.0, merging the harness it built inside Qoder into the open-source CLI and shipping it with a preview build for developers to embed.

What it means for you|Alibaba released Qwen Code 1.0, merging the harness it built inside Qoder into the open-source CLI and shipping it with a preview build for developers to embed.

One-line take|The harness, not the model, is where coding agents are won. Alibaba giving it away suggests it expects to monetise the tokens behind it.

Source: ITHome (2026-11-11)

AI SoftwareMetaX and UBTech's embodied-chip JV reported to raise ¥500 million

A joint venture between GPU maker MetaX and humanoid maker UBTech, named Xixuan Chuangzhi, is reported to have raised ¥500 million to build chips for embodied systems.

What it means for you|A joint venture between GPU maker MetaX and humanoid maker UBTech, named Xixuan Chuangzhi, is reported to have raised ¥500 million to build chips for embodied systems.

One-line take|Robot chips designed by a robot company is the vertical-integration story China's supply chain keeps reaching for. Whether the silicon is competitive is a separate question from whether the round closed.

Source: ITHome (2026-11-11)

AI SoftwareByteDance prepares a Muse-style agent

ByteDance is reportedly building a general-purpose AI agent in the mould of Meta's Muse, to be folded into the Doubao product line first.

What it means for you|ByteDance is reportedly building a general-purpose AI agent in the mould of Meta's Muse, to be folded into the Doubao product line first.

One-line take|Doubao already owns the assistant slot on hundreds of millions of phones. An agent there can transact; an agent in a browser can only browse.

Source: ITHome (2026-11-11)

AI SoftwareAnt's Lingguang 2.0 passes 3 million hosted tasks in two weeks

Ant Group said Lingguang 2.0, launched Oct 27, has hosted more than 3 million tasks and executed over 11 million individual actions. A week in, the figures were 1.16 million and 3.6 million.

What it means for you|Ant Group said Lingguang 2.0, launched Oct 27, has hosted more than 3 million tasks and executed over 11 million individual actions; a week in, the figures were 1.16 million and 3.6 million.

One-line take|Task counts are the new DAU, and they are just as easy to inflate. The number to watch is how many users come back the second week without a subsidy.

Source: Leiphone (2026-11-11)

AI SoftwareWeChat's HarmonyOS build deepens the Xiaowei assistant

WeChat for HarmonyOS opened an invite-only beta of 8.0.23.37, running Nov 10 to Dec 10, with more Xiaowei AI assistant features moving into staged rollout.

What it means for you|WeChat for HarmonyOS opened an invite-only beta of 8.0.23.37, running Nov 10 to Dec 10, with more Xiaowei AI assistant features moving into staged rollout.

One-line take|Tencent is rolling its assistant out through beta channels and grey releases, one capability at a time. For a 1.4-billion-user app, that caution is the product strategy.

Source: ITHome (2026-11-11)

AI SoftwareCoze's enterprise edition goes dark, a year after the open-source pivot

ByteDance's Coze announced on Oct 29 that its enterprise edition will shut down entirely, less than a year after the product repositioned around open source.

What it means for you|ByteDance's Coze announced on Oct 29 that its enterprise edition will shut down entirely, less than a year after the product repositioned around open source.

One-line take|Agent-building platforms are being squeezed from both ends: model vendors ship the builder for free, and enterprises would rather own the workflow. Coze was in the middle of that sandwich.

Source: Huxiu (2026-11-11)

AI SoftwareAliExpress says its AI ops agent lifted emerging-market GMV 50%

In the first seven days of Double 11, AliExpress reported average daily GMV in secondary markets up more than 50% year on year, crediting a new country-level operations agent that rewrites listings and pricing automatically.

What it means for you|In the first seven days of Double 11, AliExpress reported average daily GMV in secondary markets up more than 50% year on year, crediting a new country-level operations agent that rewrites listings and pricing automatically.

One-line take|This is the least discussed AI business model in China: not selling software, but using it to move your own inventory in markets too small to staff.

Source: Leiphone (2026-11-11)

AI SoftwareA Tsinghua-linked AI glasses startup takes seed money from an eyewear maker

Shenzhen Qingmu Digital closed a seed round from Nanjing Shike, a traditional eyewear company, with an angel round in progress. Its first product targets office work with dual waveguide displays, a microphone array, and no camera.

What it means for you|Shenzhen Qingmu Digital closed a seed round from Nanjing Shike, a traditional eyewear company, with an angel round in progress; its first product targets office work with dual waveguide displays, a microphone array, and no camera.

One-line take|No camera is a deliberate bet — it dodges the privacy fight and the battery drain. It also removes the one feature every demo reel depends on.

Source: Leiphone (2026-11-11)

AI SoftwareAlibaba moves to spin off T-Head onto the Hong Kong exchange

Alibaba said on Nov 10 that it filed to spin off chip unit T-Head as a separately listed company. T-Head submitted its A1 application confidentially on Sept 11, and the exchange has confirmed the split can proceed.

What it means for you|Alibaba said on Nov 10 that it filed to spin off chip unit T-Head as a separately listed company; T-Head submitted its A1 application confidentially on Sept 11, and the exchange has confirmed the split can proceed.

One-line take|Alibaba is separating the part of its AI story that has hard revenue from the part that has a narrative. Hong Kong gets to price the silicon on its own.

Source: TMTPost (2026-11-11)

AI SoftwareEmbodied AI funding halves as the IPO window narrows

Investors describe a sharp cooling: rounds are smaller, listings are harder to clear, and companies that raised last cycle are burning cash month by month. One fund manager called the correction a relief for operators.

What it means for you|Investors describe a sharp cooling: rounds are smaller, listings are harder to clear, and companies that raised last cycle are burning cash month by month; one fund manager called the correction a relief for operators.

One-line take|A sector where the pitch deck promises a general-purpose worker needs a capital market willing to wait a decade. That market just closed its wallet.

Source: Huxiu (2026-11-11)

AI SoftwareRobot makers start renting someone else's brain

Leju is training industrial manipulation on Ant Group's Lingbo model, and Galaxea both builds its own models and uses external ones. Foundation models are beginning to move between Chinese robotics companies.

What it means for you|Leju is training industrial manipulation on Ant Group's Lingbo model, and Galaxea both builds its own models and uses external ones, as foundation models begin to move between Chinese robotics companies.

One-line take|Outsourcing the brain compresses R&D and destroys differentiation in the same move. The robot makers who survive will be the ones whose bodies generate data nobody else can get.

Source: Huxiu (2026-11-11)

📖 The full EastVoice issue on the Chinese forum

EastVoice — five minutes a day, the global AI shift
Physix Frontier (Shenzhen) · covering foundation models, AI software, humanoid robots, autonomous driving and physical AI