Headlines · Hot List · Forum · 中文

EastVoice

PHYSIX FRONTIER EASTVOICE

2026.11.18 · Wednesday

Issue 084

FIVE MINUTES A DAY, THE GLOBAL AI SHIFT

$500 million

Agile Robots Series F round led by SoftBank

$8.5 billion

Agile Robots post-money valuation

2.4 trillion

Qwen4-Max parameters in Alibaba flagship

120 km/h

Stelato V8 L3 cross-city highway test speed

Top Stories 3 · dual commentary | Section Picks 26 · one-line takes

TOP STORIES

Unitree's Wang Xingxing says generalization, not hardware, is the hard part; Alibaba's Qwen4-Max: 2.4 trillion parameters; DeepSeek starts rationing, and users notice within two hours.

1Unitree's Wang Xingxing says generalization, not hardware, is the hard part

Speaking at the APEC CEO Summit in Shenzhen, Unitree chairman Wang Xingxing put the industry's real bottleneck on model generalization. His ask is concrete. A robot should walk into a home and start working without being retrained for that particular house.

What it means for you|Unitree's chairman says a robot should walk into a home and start working without being retrained for that particular house.

Editor's take|Every humanoid company has a demo reel. Almost none has a policy that survives a new kitchen.

Source: ITHome (2026-11-17)

2Fourier's Gu Jie draws a line at the brain

In an interview with Huxiu, Fourier Intelligence founder Gu Jie said the company will hold its boundary and stay out of building foundation models. "We don't make toys," he said. Fourier has since been linked to a Meta project, a roughly 1.5-meter, 48-kg humanoid built jointly by the two.

What it means for you|Fourier Intelligence will hold its boundary and stay out of building foundation models, per founder Gu Jie.

Editor's take|Refusing the brain is a defensible bet on being the body everyone else buys. It also caps the multiple.

Source: Huxiu (2026-11-17)

3Lei Jun: embodied AI will be bigger than cars

Xiaomi's founder told the same summit that embodied robotics is "certainly a huge industrial opportunity," with more imagination in it than the automotive business Xiaomi now depends on. He framed it against four years of rapid progress in foundation models.

What it means for you|Xiaomi's founder said embodied robotics is a huge industrial opportunity with more imagination than the automotive business Xiaomi depends on.

Editor's take|Xiaomi shipped a car with a factory and a supply chain. A humanoid is a harder promise to make from a stage.

Source: ITHome (2026-11-17)

SECTION PICKS

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

Humanoid RobotsAgile Robots raises $500M, and the revenue comes from production lines

The Munich company closed a $500 million Series F at an $8.5 billion post-money valuation, led by SoftBank with Nvidia, Hitachi and T.Capital participating. Cumulative funding now sits near $2 billion, and it bills itself as Europe's most valuable private robotics company. Chinese coverage sold it as a humanoid unicorn. TMTPost's read is that the unit actually paying the bills is factory automation.

What it means for you|Agile Robots closed a $500 million Series F at an $8.5 billion post-money valuation, with cumulative funding near $2 billion.

One-line take|A humanoid headline over an automation P&L. Eventually investors price the P&L.

Source: TMTPost (2026-11-17)

Humanoid Robots48 hours on real machines, one base model. What the embodied contest actually tested

GeekPark reports on the largest embodied-model challenge to date, where teams experimented around a single foundation model through a 48-hour final on physical robots. Both the wins and the failures were scored against hardware, not simulation.

What it means for you|Teams experimented around a single foundation model through a 48-hour final on physical robots, with wins and failures scored against hardware, not simulation.

One-line take|Simulation benchmarks are cheap. Forty-eight hours on a real arm is where the field's actual progress shows up.

Source: GeekPark (2026-11-17)

Autonomous DrivingTesla's Miami robotaxi zone nearly doubles, downtown now covered around the clock

Tesla expanded its Miami service area to include downtown and several surrounding neighborhoods, with 24-hour operation. The footprint is close to twice its previous size.

What it means for you|Tesla expanded its Miami service area to include downtown and surrounding neighborhoods, with 24-hour operation and a footprint close to twice its previous size.

One-line take|Adding square miles is easy. Adding them without a safety driver is the part regulators watch.

Source: ITHome (2026-11-17)

Autonomous DrivingStelato V8 runs China's first 120 km/h L3 cross-city highway test

Stelato, the Huawei-backed brand, says its V8 became the first vehicle in China to run an L3 highway test across cities at 120 km/h. The company frames the step from single-city trials to intercity running as a new phase for its L3 program.

What it means for you|Stelato says its V8 became the first vehicle in China to run an L3 highway test across cities at 120 km/h.

One-line take|A test permit is not a product. The interesting number will be disengagement rate per 100 km, and nobody has published one.

Source: ITHome (2026-11-17)

Autonomous DrivingXiaomi renames its driver assist and adds an iPhone key

HyperOS 1.20.0 is rolling out to Xiaomi's Pengcheng cars. The update adds an iPhone digital key and renames "end-to-end assisted driving" to "all-domain navigation assist."

What it means for you|HyperOS 1.20.0 rolling out to Xiaomi's Pengcheng cars adds an iPhone digital key and renames end-to-end assisted driving to all-domain navigation assist.

One-line take|Renaming a feature is cheaper than improving it, and it usually precedes a marketing push.

Source: ITHome (2026-11-17)

Autonomous DrivingMusk says he is coming to China, and the FSD guessing starts again

Commenting under a post about a New York summit scheduled for November 18, Musk wrote that he will be in China and a "digital version" of him would attend instead. Chinese media read it as a signal on FSD's entry.

What it means for you|Musk wrote he will be in China and a digital version of him would attend a New York summit instead; Chinese media read it as a signal on FSD's entry.

One-line take|Musk's travel schedule has been treated as a regulatory indicator for three years. It has never once been one.

Source: Chedongxi (2026-11-17)

Autonomous DrivingRobin Li: robotaxis are about 10x safer than human drivers

At the APEC summit in Shenzhen, Baidu's founder said robotaxi safety records in operation run roughly 10 times better than human drivers, and called the service one of AI's clearest deployments.

What it means for you|Baidu's founder said robotaxi safety records in operation run roughly 10 times better than human drivers.

One-line take|Per-mile crash rates favor the robot. Whether that math survives a Chinese winter and a delivery scooter is a separate question.

Source: ITHome (2026-11-17)

Autonomous DrivingA year on, is the 896-line lidar still worth the money?

Huawei's Qiankun unit launched its 896-line dual-path image-grade lidar in March, billed as the highest-spec mass-produced sensor in the world, and put it first on the AITO M9 and Maextro S800. Huxiu's owner-level reporting asks what that spec still buys a buyer twelve months later.

What it means for you|Huawei's 896-line lidar launched in March and was put first on the AITO M9 and Maextro S800; Huxiu asks what that spec still buys a buyer twelve months later.

One-line take|Sensor specs depreciate faster than any other line item on an EV. Owning the top number is a two-quarter advantage.

Source: Huxiu (2026-11-17)

Autonomous DrivingHorizon Robotics picks a public fight with Momenta

TMTPost reports that Horizon CEO Yu Kai, normally quiet, has spent weeks needling rivals. The trigger is Momenta's July IPO and its "first physical-AI stock" story, which Yu Kai has dismissed as capital-market packaging. Momenta's chip partner has since announced its own silicon.

What it means for you|Horizon CEO Yu Kai has spent weeks needling rivals, triggered by Momenta's July IPO and its first physical-AI stock story.

One-line take|Two Chinese autonomy suppliers arguing over who is more "physical" while both sell to the same three OEMs.

Source: TMTPost (2026-11-17)

World Models & Physical AIHuawei Cloud open-sources CloudRobo-WM1.2, an Ascend-native embodied world model

Released under Huawei Cloud's industry AI program, CloudRobo-WM1.2 is built natively for Ascend silicon and published as open source. It is aimed at embodied tasks rather than text.

What it means for you|Huawei Cloud released CloudRobo-WM1.2, built natively for Ascend silicon and published as open source, aimed at embodied tasks rather than text.

One-line take|Open-sourcing a world model is a hardware play. The weights are free, the training cluster is not.

Source: Xinzhiyuan (2026-11-17)

World Models & Physical AIBAAI ships PhysCapBench, a cross-domain yardstick for world models

Leiphone reports on PhysCapBench from BAAI, described as the first cross-domain benchmark for the physical fidelity of world models. It combines internet video, egocentric hand-object footage, real robot trajectories and force-tactile feedback into one scoring frame.

What it means for you|BAAI's PhysCapBench combines internet video, egocentric hand-object footage, real robot trajectories and force-tactile feedback into one scoring frame.

One-line take|A benchmark nobody adopts is a paper. A benchmark labs fear is infrastructure.

Source: Leiphone (2026-11-17)

World Models & Physical AILi Mu joins Aimotion to build industrial-grade world models

QbitAI reports that Li Mu, formerly a senior principal scientist at AWS AI, a co-creator of Apache MXNet and a CMU PhD, has taken a role at Aimotion. The startup's stated target is an industrial-grade world model.

What it means for you|Li Mu, formerly a senior principal scientist at AWS AI, has taken a role at Aimotion, whose stated target is an industrial-grade world model.

One-line take|China's world-model race just acquired a name Western labs recognize on sight.

Source: QbitAI (2026-11-17)

Foundation ModelsAlibaba's Qwen4-Max: 2.4 trillion parameters, per-task cost down nearly 60%

Released November 16, Qwen4-Max arrives with a rewritten base architecture and a per-task cost roughly 60% lower than its predecessor. GeekPark's framing is that Alibaba has moved its flagship from demo to production scheduling.

What it means for you|Qwen4-Max arrives with a rewritten base architecture and a per-task cost roughly 60% lower than its predecessor.

One-line take|Parameter counts stopped being a selling point about a year ago. Cost per completed task is the number enterprise buyers now ask for.

Source: GeekPark (2026-11-17)

Foundation ModelsSenseTime's SenseNova Loop-9B buys reasoning depth by reusing layers

Instead of scaling parameters, SenseNova Loop-9B loops a layer stack to expand effective computation and pairs it with manifold-constrained hyper-connections. The result is roughly 9B parameters that deploy inside 10 GB of VRAM, aimed at complex reasoning and agentic tasks.

What it means for you|SenseNova Loop-9B loops a layer stack to expand effective computation, yielding roughly 9B parameters that deploy inside 10 GB of VRAM.

One-line take|China's second tier has stopped chasing frontier scale and started optimizing for what actually fits on a customer's server.

Source: ITHome (2026-11-17)

Foundation ModelsCambricon ships Day-0 support for Loop-9B

Cambricon completed same-day adaptation of the Loop-9B model, covering the inference path for the looped compute architecture on its own silicon.

What it means for you|Cambricon completed same-day adaptation of the Loop-9B model, covering the inference path for the looped compute architecture on its own silicon.

One-line take|Day-0 support is the new handshake between domestic chips and domestic models. It is also the clearest sign the two camps are hardening into one stack.

Source: ITHome (2026-11-18)

Foundation ModelsKuaishou's Kling Image 4.0 goes native 8K and 2,000 characters per image

Kling Image 4.0 uses a unified understanding-and-generation architecture, supports native 8K output and can render up to 2,000 characters inside a single image. The API is open to enterprise customers and developers.

What it means for you|Kling Image 4.0 supports native 8K output and can render up to 2,000 characters inside a single image, with the API open to enterprise customers and developers.

One-line take|Rendering 2,000 legible glyphs is a document problem, not an art problem. That is where the money is.

Source: ITHome (2026-11-17)

Foundation ModelsmiHoYo's Gamer model plays Honkai: Star Rail on its own

miHoYo's Honkai AI team published Gamer, a general vision-language-action foundation model that operates games through the same input a human uses, covering both 3D and 2D titles.

What it means for you|miHoYo's Honkai AI team published Gamer, a vision-language-action foundation model that operates games through the same input a human uses, covering 3D and 2D titles.

One-line take|Games are the cheapest sim-to-real rig ever built. A studio with its own engine realized that before most robotics labs did.

Source: ITHome (2026-11-17)

Foundation ModelsChina's flagship models quietly cap their own cyber-offense ability

Bloomberg reports that the newest models from Alibaba, Tencent and Moonshot all restrict offensive cyber capability, while strengthening vulnerability detection and defense. The shift tracks tighter AI technology export controls.

What it means for you|Bloomberg reports the newest models from Alibaba, Tencent and Moonshot all restrict offensive cyber capability while strengthening vulnerability detection and defense.

One-line take|Every US lab is being sued over agent behavior this month. Chinese labs are self-capping in advance. Different liability regimes, different reflexes.

Source: ITHome (2026-11-17)

Foundation ModelsRobin Li warns of an AI bubble

In the same Shenzhen speech, Baidu's founder said compute and infrastructure spending is enormous while results at the application layer have underdelivered. If the app layer does not break through, he said, bubble risk becomes visible.

What it means for you|Baidu's founder said compute and infrastructure spending is enormous while results at the application layer have underdelivered, making bubble risk visible without a breakthrough.

One-line take|The CEO whose company monetizes AI cloud is warning about AI capex. Read it as a forecast about who gets funded next year.

Source: ITHome (2026-11-17)

Foundation ModelsTencent's Hy4 Flash costs a twentieth of Hy4 per task

Released November 16, Hy4 Flash keeps Hy4's vision and agent capabilities while cutting latency and cost. GeekPark measured per-task cost at about one twentieth of the full model.

What it means for you|Hy4 Flash keeps Hy4's vision and agent capabilities while cutting latency and cost, with per-task cost about one twentieth of the full model.

One-line take|"Good enough and cheap" is the product most Chinese enterprises were actually waiting for.

Source: GeekPark (2026-11-17)

AI SoftwareDeepSeek starts rationing, and users notice within two hours

DeepSeek told users that long conversations idle for more than seven days would stop being available, capped single inputs at roughly 40,000 characters and introduced daily quotas, citing compute limits. Some users hit the ceiling in under two hours.

What it means for you|DeepSeek said long conversations idle more than seven days would stop being available, capped single inputs at roughly 40,000 characters and introduced daily quotas, citing compute limits.

One-line take|Free inference at this scale was always a subsidy. The question is whether Chinese consumers will pay, and the first data point arrives this quarter.

Source: TMTPost (2026-11-17)

AI SoftwareBaidu Q3: AI is 53% of core revenue, and the stock still fell

Total revenue fell 1% to ¥30.807 billion. AI new business brought in ¥13.2 billion, up 37%, or 53% of core revenue, with cloud infrastructure at ¥8.2 billion of it. Investors marked the stock down anyway.

What it means for you|Baidu's total revenue fell 1% to ¥30.807 billion while AI new business brought in ¥13.2 billion, up 37%, or 53% of core revenue.

One-line take|More than half of core revenue now comes from AI, and the market's response was a selloff. Growth quality, not growth share, is what is being priced.

Source: Huxiu (2026-11-17)

AI SoftwareQingtianzu closes a nine-figure strategic round to rent out robots

The robot rental platform, co-founded by Agibot and Feikuo, closed a strategic round in the hundreds of millions of RMB, with investors bringing regional industry access, entertainment IP and logistics scenarios. Its stated thesis is that leasing is only the entry point, and the real revenue is robots doing work.

What it means for you|Qingtianzu closed a strategic round in the hundreds of millions of RMB, with investors bringing regional industry access, entertainment IP and logistics scenarios.

One-line take|If the money is in the work, the lessor is a staffing agency with extra steps. That may be the most honest humanoid business model on the board.

Source: Huxiu (2026-11-17)

AI SoftwareLi Jiaqi sold tokens on a livestream, 100,000 subscriptions in 30 minutes

On November 2 the top livestreamer put a Qwen office membership card on air. 100,000 monthly cards sold in half an hour, roughly 2 trillion tokens of capacity moving through a shopping channel.

What it means for you|Li Jiaqi put a Qwen office membership card on air; 100,000 monthly cards sold in half an hour, roughly 2 trillion tokens of capacity through a shopping channel.

One-line take|Tokens are being merchandised like skincare. Whatever the unit economics, distribution has already been solved by people who sell lipstick.

Source: Huxiu (2026-11-17)

AI SoftwareApple adds an AIGC filing gate to App Store apps in China

App Store Connect now requires a declaration of AI-generated content services, filing information and proof of completed security assessment for apps distributed in mainland China. Existing apps have until January 31, 2027.

What it means for you|App Store Connect now requires a declaration of AI-generated content services, filing information and proof of completed security assessment for apps in mainland China, with existing apps due by January 31, 2027.

One-line take|Compliance as a growth filter. Small AI app shops will lose two months to paperwork and larger ones will absorb it as a line item.

Source: ITHome (2026-11-17)

AI SoftwareMarsh: attackers are already using AI agents to hunt for vulnerabilities

Reuters reports that insurance broker Marsh has seen signs of attackers using AI tools and agents to find security holes faster and at larger scale, with corporate cyber incident reports rising in step.

What it means for you|Reuters reports Marsh has seen signs of attackers using AI tools and agents to find security holes faster and at larger scale, with corporate cyber incident reports rising in step.

One-line take|The offense got agents first. Insurance premiums are where that asymmetry gets priced.

Source: ITHome (2026-11-17)

📖 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