Headlines · Hot List · Forum · 中文

EastVoice

PHYSIX FRONTIER EASTVOICE

2026.10.24 · Saturday

Issue 067

FIVE MINUTES A DAY, THE GLOBAL AI SHIFT

¥1.699 billion (about $239 million)

Unitree Technology 2025 revenue at 60.44% gross margin

¥100 billion (about $15 billion)

DeepSeek second private round with CATL and Tencent

428-billion-parameter

MiniMax M3.1 Flash open-source weights released

312

China embodied-AI funding deals in first half 2026

Top Stories 3 · dual commentary | Section Picks 24 · one-line takes

TOP STORIES

Unitree's 60% gross margin meets its first robot price war, DeepSeek closes a second round of about ¥100 billion, and DeepSeek and MiniMax clear China's new export rules.

1Unitree's 60% gross margin meets its first robot price war

Unitree Technology booked ¥1.699 billion (about $239 million) in 2025 revenue at a 60.44% gross margin, roughly 16 points above its closest Chinese peers. UBTECH sat at 37.67%, Dobot at 46.49% and DEEP Robotics at 52.83%.

Editor's take|A 16-point margin gap is either real pricing power or a subsidy that hasn't surfaced yet. A price war will tell you which.

Source: Huxiu (2026-10-23)

2Humanoid rental prices fall 70% while vendors keep signing channels

A full-package daily rental for a Unitree U2 with a technician has slid from ¥10,000 in early 2025 to under ¥3,000 by October 2026. Vendors are still expanding their dealer networks.

What it means for you|A full-package daily rental for a Unitree U2 with a technician has slid from ¥10,000 in early 2025 to under ¥3,000 by October 2026, while vendors are still expanding their dealer networks.

Editor's take|When rentals drop that fast the hardware is a commodity. Whoever locks up the deployment channel wins the next year, margin or not.

Source: TMTPost (2026-10-23)

3Wang Xingxing tells the industry to brace for two hard years

At Unitree's Physical AI partner conference on Oct 22, with more than 3,000 guests, Wang Xingxing said embodied robots still lack the kind of general model that lifted GPT-6. The title of his talk was "Trust AI, as much as you can."

Editor's take|The founder of China's hottest robotics stock telling buyers to wait is a rare honest signal. He wouldn't say it if demand were roaring.

Source: Huxiu (2026-10-23)

SECTION PICKS

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

Humanoid RobotsEmbodied-AI investors are already underwater

China recorded 312 embodied-AI funding deals in the first half of 2026, more than the 227 closed by foundation-model startups, according to IT Juzi. Early backers of several now-listed names are showing mark-to-market losses.

What it means for you|China recorded 312 embodied-AI funding deals in the first half of 2026, more than the 227 closed by foundation-model startups, and early backers of several now-listed names are showing mark-to-market losses.

One-line take|Deal count beating foundation models is a sign of money chasing demo reels. The losses show up later, on the secondary market.

Source: Huxiu (2026-10-23)

Autonomous DrivingBYD's new Qin PLUS arrives Oct 27 with a roof-mounted lidar

BYD set the launch of its refreshed Qin PLUS for Oct 27. The sedan gains a roof lidar, a 2,800 mm wheelbase and a larger body, with a new design language.

What it means for you|BYD set the launch of its refreshed Qin PLUS for Oct 27, and the sedan gains a roof lidar, a 2,800 mm wheelbase and a larger body, with a new design language.

One-line take|Lidar on a mass-market sedan means the sensor war is effectively over. The cost curve did what years of policy debate couldn't.

Source: ITHome (2026-10-23)

Autonomous DrivingAudi may build its flagship A8 in China, with Huawei's help

German outlet Manager Magazin reported that Audi is in talks with Huawei to co-develop its A8 sedan for the Chinese market, under pressure from global CEO Gernot Döllner to move faster.

One-line take|A German flagship handing its software stack to a Chinese vendor is the clearest admission yet of who leads in smart cockpits.

Source: Huxiu (2026-10-23)

Autonomous DrivingBYD expects L3 autonomy to reach the road by 2028

BYD said it expects conditional L3 autonomy to start reaching production vehicles in China around 2028, framing it as the next step beyond today's assisted-driving systems.

What it means for you|BYD said it expects conditional L3 autonomy to start reaching production vehicles in China around 2028, framing it as the next step beyond today's assisted-driving systems.

One-line take|2028 keeps slipping right, for everyone. The liability framework, not the sensor suite, is what's setting the date.

Source: iFanr (2026-10-23)

Autonomous DrivingXpeng's MONA L05 launch turns into a marketing spat

Chinese commentators questioned Xpeng's Oct 15 launch campaign for the MONA L05, arguing the company leaned on attention-grabbing marketing over product substance.

One-line take|When a carmaker's launch story is the marketing itself, buyers notice. The product still has to survive the test drive.

Source: Huxiu (2026-10-23)

Autonomous DrivingAvatr ships a smart-cockpit update with dual-map navigation

Avatr began pushing AVATR.OS 6.1.0 to vehicles on Oct 23, adding dual-map switching, a "stay powered after exit" mode and an interactive digital pet.

What it means for you|Avatr began pushing AVATR.OS 6.1.0 to vehicles on Oct 23, adding dual-map switching, a "stay powered after exit" mode and an interactive digital pet.

One-line take|OTA feature races are now about retention, not capability. A pet and a stay-powered mode are aimed at daily habit, not benchmarks.

Source: ITHome (2026-10-23)

World Models & Physical AIA single RTX 5090 runs a 34B real-time world model

A Chinese team demoed a 34-billion-parameter real-time world model running on one RTX 5090, turning a room into a grassland and adding snow on command. It won an award at Nvidia's startup showcase in Suzhou on Oct 15.

One-line take|A world model that needs a rack of GPUs is a paper. One that runs on a consumer card is a product.

Source: QbitAI (2026-10-23)

World Models & Physical AIAlibaba's AMap releases λ0, a whole-body robot foundation model

Alibaba's AMap published λ0, described as the first whole-body foundation model built on open-world human-action data. The robot learns locomotion, posture and two-arm motion from human movement alone.

One-line take|Training on human motion instead of robot teleoperation slashes the data bill. The open question is how well it transfers to hardware with different joints.

Source: QbitAI (2026-10-23)

World Models & Physical AIWorld models find a paying customer in gaming and social

Chizicheng worked with video-generation startup Viggle to ship 3D mini-games built on a world model, one of the first commercial deployments of the technology outside research labs.

One-line take|Games are the cheapest place to prove a world model keeps its own rules. If it holds there, industrial simulation is next.

Source: AI Era (2026-10-23)

World Models & Physical AINanjing bets on physical AI with 50 deployment scenarios

At a Nanjing physical-AI event, three Turing Award laureates spoke and the city unveiled 50 application scenarios, positioning itself as a live testbed for embodied systems.

One-line take|Cities competing on robot scenarios is China's version of a regulatory sandbox. The city that amasses the most deployment data seeds the next cohort.

Source: QbitAI (2026-10-23)

Foundation ModelsDeepSeek closes a second round of about ¥100 billion

DeepSeek completed a second private round raising roughly ¥100 billion (about $15 billion), with CATL and Tencent among the investors, per the Wall Street Journal. Demand pushed the raise to nearly twice its original target.

What it means for you|DeepSeek completed a second private round raising roughly ¥100 billion (about $15 billion), with CATL and Tencent among the investors, per the Wall Street Journal, and demand pushed the raise to nearly twice its original target.

One-line take|A round this size, that oversubscribed, says Chinese institutional money stopped waiting for a homegrown OpenAI to appear on its own.

Source: ITHome (2026-10-23)

Foundation ModelsDeepSeek and MiniMax clear China's new export rules

Both labs completed registration to open-source and license their models overseas, days after Beijing published draft restrictions covering 23 AI technologies. Anything already in the public domain through open-source release or publication is exempt.

What it means for you|Both labs completed registration to open-source and license their models overseas, days after Beijing published draft restrictions covering 23 AI technologies, and anything already in the public domain through open-source release or publication is exempt.

One-line take|The public-domain carve-out does exactly what the labs hoped. Publish the weights first, and the export rule never reaches you.

Source: ITHome (2026-10-23)

Foundation ModelsKimi K3.1 teaser drops with no launch date

Moonshot AI's official account posted a 24-second teaser showing Kimi Code running on an old desktop, with "Model: Kimi K3.1" in the corner. No release date was announced.

One-line take|Teasers ahead of a launch are standard now. The interesting artifact is the model card, not the video.

Source: ITHome (2026-10-23)

Foundation ModelsLeak claims Kimi K3.1 generates 3D assets at one-fourteenth the cost

A leaked comparison puts Kimi K3.1's industrial-grade interactive 3D asset generation at 1/14 the cost of GPT-6 Astra, delivering in about 15 minutes.

What it means for you|A leaked comparison puts Kimi K3.1's industrial-grade interactive 3D asset generation at 1/14 the cost of GPT-6 Astra, delivering in about 15 minutes.

One-line take|Cost per asset is the only number a 3D studio cares about. A 14x gap would relocate work faster than any benchmark chart.

Source: Leiphone (2026-10-23)

Foundation ModelsMiniMax open-sources 428B-parameter M3.1 Flash weights

MiniMax released the weights for M3.1 Flash, a 428-billion-parameter model, under an open license.

One-line take|428B open weights are a serving problem as much as a model. Most teams will call it through an API, not run it on their own racks.

Source: ITHome (2026-10-23)

Foundation ModelsByteDance's Volcengine ships Seedance 2.5 4K for film-grade video

Volcengine, ByteDance's cloud arm, released Seedance 2.5 4K, trained more heavily on facial expression, fine motion and image detail for film-grade generation.

One-line take|Video models keep racing on resolution because it's the headline. The studios buying them care more about shot-to-shot consistency.

Source: ITHome (2026-10-23)

Foundation ModelsChina's token prices doubled in a year, and so did throughput

A CAICT report found the average price of Chinese model tokens roughly doubled in 2026, while per-second throughput also doubled. Volume, not discounting, is driving the bill.

What it means for you|A CAICT report found the average price of Chinese model tokens roughly doubled in 2026, while per-second throughput also doubled, and volume, not discounting, is driving the bill.

One-line take|Cheap-per-token stopped being the pitch. Buyers now pay for speed and reliability, which is a healthier market to sell into.

Source: ITHome (2026-10-23)

AI SoftwareAlipay lands Tesla, completing a trio of Chinese super-apps

Alipay became the third Chinese giant wired into Tesla vehicles in China, after Tencent connected WeChat and ByteDance put Doubao on the car's system. Alipay joined on Oct 20.

What it means for you|Alipay became the third Chinese giant wired into Tesla vehicles in China, after Tencent connected WeChat and ByteDance put Doubao on the car's system, and Alipay joined on Oct 20.

One-line take|Tesla's China strategy is now a set of distribution deals with local super-apps. The car is becoming a storefront with wheels.

Source: Huxiu (2026-10-23)

AI SoftwareQianwen Go! lets users build and share their own apps

Alibaba's Qianwen Go! launched a "My Apps" feature on Oct 23, letting users build shareable, long-running apps by describing them in plain language. Output persists and updates its own data.

What it means for you|Alibaba's Qianwen Go! launched a "My Apps" feature on Oct 23, letting users build shareable, long-running apps by describing them in plain language, and output persists and updates its own data.

One-line take|"Describe an app, get an app" is table stakes now. Persistence and sharing are what turn a demo into a platform.

Source: Leiphone (2026-10-23)

AI SoftwareMeituan hands every restaurant a free AI store manager

At its 2026 delivery conference on Oct 22, Meituan launched an AI "store manager" for food merchants and said it will give it to all restaurants nationwide. Some trial merchants grew revenue more than 30%.

What it means for you|At its 2026 delivery conference on Oct 22, Meituan launched an AI "store manager" for food merchants and said it will give it to all restaurants nationwide, and some trial merchants grew revenue more than 30%.

One-line take|Giving the tool away free is how Meituan locks in merchant data. The AI is the hook; the ordering flow is the business.

Source: Leiphone (2026-10-23)

AI SoftwareAgibot's parent has made 47 investments in two years

Zhiyuan (Agibot) appeared as an investor in 47 deals between August 2024 and October 2026, most of them into its own robot supply chain, including a recent pre-A round in a chip startup.

One-line take|A robot maker funding its own suppliers is vertical integration by another name. It caps costs today and its own sourcing options later.

Source: TMTPost (2026-10-23)

AI SoftwareTencent's QClaw shuts down after 188 days

Tencent said QClaw will stop operating on Dec 24, 2026 due to business adjustments, with a migration path to WorkBuddy. The product ran for roughly 188 days.

What it means for you|Tencent said QClaw will stop operating on Dec 24, 2026 due to business adjustments, with a migration path to WorkBuddy, and the product ran for roughly 188 days.

One-line take|Personal assistants are being repriced every quarter. Tencent would rather merge two products than keep funding both.

Source: Huxiu (2026-10-23)

AI SoftwareWho is betting on Personal AI, and who is afraid of it

China's Personal Agent field got crowded in weeks: Alibaba's Qianwen Go, a standalone Tencent agent, ByteDance's system-level agent, StepFun's agent phone, Ant's Lingguang, and Today AI raised more than $200 million.

One-line take|Six well-funded entrants in one month means the category is past the interesting stage. Distribution decides it now, not model quality.

Source: TMTPost (2026-10-23)

AI SoftwareChinese AI short dramas go overseas and hit a margin wall

AI-generated short dramas expanded overseas quickly in 2026, but producers say revenue has not kept pace with output. Some compare it to China's own domestic land-grab a year earlier.

What it means for you|AI-generated short dramas expanded overseas quickly in 2026, but producers say revenue has not kept pace with output, and some compare it to China's own domestic land-grab a year earlier.

One-line take|Cheaper production multiplies supply faster than it grows demand. The oversupply that hit domestic platforms is now exporting itself.

Source: Huxiu (2026-10-23)

📖 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