EastVoice

PHYSIX FRONTIER EASTVOICE

2026.08.25 · Tuesday

Issue 019

FIVE MINUTES A DAY, THE GLOBAL AI SHIFT

2,056 robots from 666 teams

Competitors at the 2nd World Humanoid Robot Games

$900M+

XPeng raised for its robot business valuation

30-second video generation

Alibaba Cloud's Wan3.0 model capability

15 tumors doctors missed

Caught by Alibaba DAMO's liver-cancer model

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

TOP STORIES

Humanoid Robot Games achievements, Xiaomi's AI SoC launch, and XPeng's massive robotics funding round.

1Liang Wenfeng said multimodality "isn't the main line." Then DeepSeek shipped a vision model.

In an earlier investor meeting, DeepSeek's founder argued multimodality is "a component, not the main line" for intelligence. The company has now shipped native multimodal support in the V4 series, per TMTPost's report.

Editor's take|Saying one thing and shipping the other is fine when the shipping is this good. Actions beat positioning every time.

Source: TMTPost (2026-08-24)

2GLM-5.3-Max debuts at No. 13 on the blind-test ladder; Kimi K3-Max cracks the top 10

Arena's Week 34 ranking (Aug 17–23) shows Zhipu's GLM-5.3-Max entering at No. 13 with an ELO of 1487, the strongest debut of the period, while Moonshot AI's Kimi K3-Max broke into the top 10.

Editor's take|A top-15 debut used to be a launch event. Now it's a Tuesday.

Source: ITHome (2026-08-24)

3Moonshot retires Kimi K2.5, its first trillion-parameter multimodal model

Moonshot AI announced its first-generation trillion-parameter multimodal model, Kimi K2.5, will end service at the end of this month, about seven months after its January open-source release.

Editor's take|Trillion-parameter models now have a product lifecycle. That is progress, and also a warning about how fast this stack churns.

Source: ITHome (2026-08-24)

SECTION PICKS

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

AI Foundation ModelsAlibaba Cloud's Wan3.0 goes live: 30-second clips, document input, "real-world fidelity"

Alibaba Cloud's video model Wan3.0 is now generally available, generating up to 30 seconds of video per run and accepting document inputs (doc, xls, ppt, pdf, md) for the first time.

What it means for you|Wan3.0 accepts document inputs like doc, xls, ppt, pdf, md for the first time.

One-line take|The video-model wars moved from "looks good" to "fits your workflow." Wan just joined the office.

Source: ITHome (2026-08-24)

AI Foundation ModelsAlibaba DAMO's liver-cancer model caught 15 tumors doctors missed

DAMO Academy's DAMO LiON detects liver lesions on CT scans down to 1 cm. In a two-month prospective trial it flagged 15 malignant tumors that had been overlooked, most around 1 cm.

What it means for you|The model flagged fifteen malignant tumors that had been overlooked by doctors in a trial.

One-line take|Fifteen tumors found late would have been fifteen different outcomes. AI radiology's pitch was never accuracy, it was second chances.

Source: ITHome (2026-08-24)

AI Foundation ModelsMoore Threads pitches Prefill-as-a-Service to attack long-context costs

Moore Threads published a technical whitepaper proposing a Prefill-as-a-Service paradigm on its MTT S5000 accelerator, targeting long-context workloads like AI agents, code generation and long-document analysis.

One-line take|Selling prefill as a service is how GPU makers hedge against the day inference gets commoditized. Smart hedge.

Source: ITHome (2026-08-24)

AI Foundation ModelsSenseTime Medical's $100M+ round values it past ¥10B, with Pre-IPO chatter

SenseTime Medical closed a B-round of over $100M in July, pushing its post-money valuation past ¥10 billion (~$1.4B), on a "medical world model" pitch, with reports it has entered Pre-IPO preparation.

One-line take|"Medical world model" is a hard sell to hospitals and an easy sell to VCs. The market will decide which one matters.

Source: TMTPost (2026-08-24)

AI Foundation ModelsMystery model Ox Alpha: are GLM fingerprints all over it?

China's Edge AI Daily newsletter highlights researcher Ben Davis's claim, at "99% confidence," that the anonymous Ox Alpha model carries technical fingerprints of Zhipu's GLM-5.x family, though GLM-5.3's text-only status leaves a gap. Microsoft's MAI is the other suspect.

One-line take|An anonymous model that out-scores named ones is the best marketing nobody paid for. Anonymity is the new launch strategy.

Source: TMTPost (2026-08-24)

AI Software & ApplicationsXiaomi's Surge O3: a 24-billion-transistor AI flagship SoC

Xiaomi unveiled Surge O3, its first AI flagship SoC packing 24 billion transistors, the centerpiece of an Aug 24 self-design push that also included a battery chip and a charging chip.

One-line take|Three self-designed chips in one launch means Xiaomi is done renting its silicon roadmap.

Source: ITHome (2026-08-24)

AI Software & ApplicationsXiaomi VP: chip R&D buys "a ticket to the next decade," not a money pit

Xiaomi group VP Zhu Dan told CCTV Finance that in the AI era, compute underlies every smart experience, and without chip capability you cannot define products from the bottom up. The investment, he said, buys "a ticket to the next decade."

One-line take|"Not burning money, buying a ticket" is the most honest thing a phone-maker has said about chip R&D in years.

Source: ITHome (2026-08-24)

AI Software & ApplicationsWeifan Intelligence banks a ¥100M+ seed-plus round for robot edge-AI chips

Beijing's Weifan Intelligence closed a seed-plus round above ¥100M (~$14M), backed by repeat investors plus CICC Capital, Nine-Kun Venture and humanoid maker Songyan Power, to push robot-side AI chips and its BrainOS stack toward production.

One-line take|A robot maker investing in a chip startup is a supply-chain vote, not a financial one. Watch who piles in next.

Source: TMTPost (2026-08-24)

AI Software & ApplicationsXPeng's robot arm raises $900M+ at a $6.3B valuation

XPeng's robotics business closed its first external round at over $900M, a post-money valuation of $6.3B (~¥43B), a record single private round for China's embodied-AI sector. CEO He Xiaopeng promises the first high-end humanoid will be "more human-like, safer, no remote control."

One-line take|$6.3B for a robotics side-project — the market just told carmakers their next act matters more than their current one.

Source: ITHome (2026-08-24)

AI Software & ApplicationsAlibaba raises HK$80B for AI as cloud growth picks up

Alibaba reported quarterly revenue of ¥268.95B, up 9% YoY, with cloud growth accelerating and spending climbing, and it is raising HK$80B (~$10.2B) earmarked for AI.

One-line take|Letting cloud growth fund the AI war chest is the cleanest balance-sheet story in Chinese Big Tech right now.

Source: TMTPost (2026-08-24)

AI Software & ApplicationsA "cyber matchmaker" raises ¥15M, and AI's gold rush reaches dating

A Chinese AI-powered matchmaking startup closed a ¥15M angel round, the latest sign that putting "AI" on a business plan shortens the fundraising road.

One-line take|When AI meets matchmaking, the algorithm's incentives deserve a prenup.

Source: TMTPost (2026-08-24)

AI Software & ApplicationsA former security guard reached the AI finals; a high-schooler banked ¥250K

At an AI contest in Shanghai, 24-year-old Peng Mingyu, a former security guard, made it to the finals, while a high-school student took home a ¥250,000 prize.

One-line take|The fastest-growing AI adopters are no longer engineers. That is either inspiring or terrifying, depending on who you are.

Source: QbitAI (2026-08-24)

Humanoid RobotsA robot takes the first weightlifting gold of the Games, lifting 16 kg

At the 2nd World Humanoid Robot Games in Beijing, the Beijing Renxing-Huake team's robot lifted 16 kg for the weightlifting crown, stepping up repeatedly from the 15 kg base weight.

One-line take|From dancing to deadlifts, the competition floor is turning into a benchmark lab.

Source: ITHome (2026-08-24)

Humanoid RobotsZhiyuan's A3 takes tai-chi gold; medal tally now 3 gold, 3 silver, 2 bronze

Zhiyuan Robotics' Expedition A3 won the martial-arts tai-chi gold at the Games, lifting the team to 3 gold, 3 silver and 2 bronze, top of the medal table.

One-line take|The motion-control showcase has officially become a medal race. Everyone is training for next year already.

Source: ITHome (2026-08-24)

Humanoid RobotsLong-jump gold at 7.97 m, on three improving attempts

The Tianjiao team's robot took long-jump gold with 7.97 m, improving across three tries (7.32, 7.62, 7.97), ahead of Tianshuo's 7.24 m and Tiangong's 7.00 m.

One-line take|Three successive personal bests is a control-loop story, not a hardware story.

Source: ITHome (2026-08-24)

Humanoid RobotsTiangong's 400 m win came with a face-covering finish, and no one taught it that

Tiangong Omni won the 400 m small-class final in 45.66 s, finishing with a viral "face-covering" run. Its development team insists the gesture was not choreographed, saying the robot came up with it on its own.

One-line take|The robot invented its own meme. Emergent behavior, broadcast live.

Source: ITHome (2026-08-24)

Humanoid RobotsFAW and Leapmotor deepen ties across EVs, ADAS and embodied AI

FAW and Leapmotor signed a strategic cooperation agreement in Changchun covering capital, NEVs, intelligent driving, powertrains, batteries, chassis, lightweight parts and embodied-intelligence robots.

One-line take|Two carmakers signing a robot clause into their agreement tells you where the industry thinks the margin lives next.

Source: ITHome (2026-08-24)

Humanoid RobotsLeapmotor confirms it has plans in embodied AI, details "coming soon"

At its mid-year results call, Leapmotor said NEV makers with full-stack in-house R&D are well positioned to build embodied AI, confirmed the company has plans in the field, and promised announcements soon.

One-line take|Every Chinese EV maker now has a robot strategy. The ones saying "just rumors" are the ones to watch.

Source: ITHome (2026-08-24)

Humanoid RobotsJD.com's "Robot Home" runs full-time support for Games robots

As an official partner of the Games, JD.com set up a "Robot Home" service center providing full-event support for the 2,056 competing robots from 666 teams across 16 countries.

One-line take|Selling after-sales service to robots, now that is a business model.

Source: Leiphone (2026-08-24)

Humanoid RobotsWRC 2026: the robots' "probation period" is over

A review from the World Robot Conference floor argues robots have proven they can work, stacking parcels, brewing coffee, folding laundry. The question now is scale, not capability.

One-line take|When the demos stop being the story, the industry has actually arrived.

Source: TMTPost (2026-08-24)

Autonomous DrivingQianli Technology's H1: revenue ¥5.12B, ADAS on 16 models, net profit up 151%

Qianli Technology reported H1 2026 revenue of ¥5.12B, up 22.4% YoY, and net profit of ¥78M, up 151%, with its intelligent-driving business covering 16 vehicle models.

One-line take|Margins are thin but the curve is right. ADAS suppliers graduate one quarter at a time.

Source: ITHome (2026-08-24)

Autonomous DrivingChengdu Auto Show: everyone is prepping for L3 before the rulebook is final

At the Chengdu Auto Show, Chinese brands showcased L3-ready architectures and in-car AI agents ahead of the mandatory autonomous-driving standard taking effect in 2027. HIMA touted its Xiangjie G9 as a "first L3-architecture" SUV with a 120 km/h L3 test license.

One-line take|Building for L3 before the rulebook is final is a classic Chinese speed-run, and it usually works.

Source: Baidu Baijiahao (2026-08-21)

World Models / Physical AIShengshu Tech lays out a five-tier roadmap to general world models

At WRC 2026, Shengshu Technology's co-founder and chief scientist Zhu Jun presented a five-level development roadmap for general world models, from large-model foundations toward embodied physical-world understanding.

One-line take|Roadmaps matter less than release cadence. The field will grade Shengshu on shipping, not slides.

Source: Leiphone (2026-08-24)

World Models / Physical AINative omni-modal world models: from simulation to interaction

At WRC 2026's "Physical AI Leaders Forum," researchers argued the frontier is shifting from generative simulation to interactive, native omni-modal world models that pair perception with action.

One-line take|"Video is imagination, interaction is understanding." The field is finally saying that out loud.

Source: QbitAI (2026-08-24)

World Models / Physical AIOrbbec teams with Liqing to feed world models better 3D data

During WRC, Orbbec signed a strategic partnership with Liqing Intelligence to improve world-model training data, covering high-quality 3D depth data and visual-sensor-to-simulation alignment.

One-line take|World models are only as good as the data they eat. The sensor guys just got a seat at the table.

Source: 10jqka (2026-08-24)

📖 The full EastVoice issue on the Chinese forum

📈 Physix Frontier · Alpha — the investor's read

Headline teardowns · curated signals · market moves. Five minutes to find the alpha of the AI era →

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