EastVoice · AI · 2026-08-22 · Issue #44

Editor’s Note: China’s AI week peaked on the trading floor. Unitree surged as much as 629% intraday to a ¥444.9B peak market cap on listing day, while SenseTime, Ant Group and Alibaba all pushed out open-weight releases within 48 hours. Also on the board: JPMorgan’s bull case on Alibaba Cloud, a first-of-its-kind embodied-model filing, and an autonomous light truck that hauls 4.2 tons.

Commentary: The pattern is getting hard to miss. Chinese labs now ship open models the way US labs ship press releases, and this year’s World Robot Conference stopped being a showroom and turned into a sourcing floor. Watch the unit economics, not the demo reel.

---

I. AI Foundation Models

1. SenseTime Opens Up SenseNova U1.5 Lite, an 8B Multimodal Model Built for Real Visual Workflows

   • Summary: SenseTime (商汤科技) released the stable version of SenseNova U1.5 Lite, an open-source 8B natively unified multimodal model. It handles 3-4K context, juggles subject, quantity, spatial-relation, typography and layout constraints at once, outputs native 4K images, and was retrained around real visual-production tasks.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: Open-sourcing an 8B model that does production-grade 4K generation is a quiet shot at the mid-tier closed models. Small, cheap and controllable is where the market is heading.

2. Ant Group Open-Sources Ling-3.0-tiny/flash Base, Training Checkpoints Included

   • Summary: Ant Group (蚂蚁集团) released the Ling-3.0-tiny-base and Ling-3.0-flash-base checkpoints and, unusually, is also publishing key training-process milestones so outsiders can trace how the models were built.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: Opening up intermediate checkpoints is a transparency move most labs won’t touch. Ant is buying developer trust with receipts.

3. Alibaba Cloud Adds GLM-5.3 and DeepSeek-V4-Pro to Its Qwen Platform

   • Summary: Alibaba Cloud (阿里云) plugged GLM-5.3 and DeepSeek-V4-Pro stable versions into its Qianwen AI platform, with APIs live for developers and both pitched at coding and agent workloads.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: The country’s biggest model marketplace now openly resells its rivals’ flagships. Aggregation is becoming Alibaba’s moat, not any single model.

4. DeepSeek Harness Hits Third Release in a Week, Turns Claude Code and Codex Into Sub-Agents

   • Summary: DeepSeek shipped rc.8 of its open-source agent workbench Harness just two days after rc.7. The August 13 debut passed 30,000 GitHub stars in its first day; the new version lets Harness orchestrate Claude Code and Codex as subordinate agents.
   • Source: TMTPost · 2026-08-21
   • Editor’s Take: DeepSeek’s endgame may not be another model, it could be the harness everyone’s agents run inside. Control the orchestration layer and you tax the whole stack.

5. JPMorgan: Alibaba Cloud’s 12% Margin Is Systematically Undervalued

   • Summary: JPMorgan analysts argue Alibaba Cloud’s current 12% operating margin understates maturity, with steady-state ROIC closer to 20% once recently deployed GPUs and data centers finish ramping from around 60% utilization.
   • Source: Leiphone · 2026-08-21
   • Editor’s Take: The vintage-stacking math says Alibaba’s CapEx binge converts into profit if utilization keeps climbing. That is the whole China AI capex trade in one chart.

6. UBTech’s Embodied Model Is One of the First to Clear China’s GenAI Filing

   • Summary: UBTech Robotics’ (优必选) Walker embodied-intelligence model passed the CAC review and completed China’s generative-AI service filing, among the first embodied models to clear the process.
   • Source: QbitAI · 2026-08-21
   • Editor’s Take: A compliance stamp matters more than a benchmark score when selling robots to Chinese enterprises. UBTech just got the paperwork that unlocks procurement.

7. Lingxi Zhiyong Attacks the “Model-Only” Myth With a Model + Harness Stack for Factories

   • Summary: USTC robotics PhD Duan Yifan’s Shanghai startup Lingxi Zhiyong (灵犀智涌) argues pretrained models alone fail on real production lines, where deployment success runs under 60%, and sells a model-plus-harness architecture instead.
   • Source: Leiphone · 2026-08-21
   • Editor’s Take: The bottleneck in industrial embodied AI is wiring, not weights. Anyone who treats the factory as a pure model problem will keep losing the last mile.

8. “Compute Loans” Arrive in Guangzhou, Making GPU Usage the New Collateral

   • Summary: Chinese media report banks in Guangzhou now factor companies’ AI compute consumption into lending decisions, while Wall Street packages GPU assets for pension funds. Compute is being securitized as an asset class on both shores.
   • Source: TMTPost · 2026-08-21
   • Editor’s Take: When banks underwrite GPU rental contracts, compute has officially become real estate. The question shifts from “is it overpriced” to “who’s the landlord”.

---

II. AI Software & Applications

1. MiningLamp and Hikrobot Bring “Agent + Embodied” Robots to Commercial Scenes at WRC

   • Summary: MiningLamp Technology (明略科技, 2718.HK) and Hikrobot (海康机器人) co-exhibited at WRC 2026, pairing AI agents with embodied machines for commercial service deployments.
   • Source: QbitAI · 2026-08-21
   • Editor’s Take: The smart move at robot shows this year is software-plus-hardware pairing. Agents that never move and robots that never reason both lose deals.

2. YMTC Files for a STAR Market IPO, Seeking ¥33 Billion

   • Summary: Yangtze Memory (长江存储), China’s NAND champion, had its STAR Market IPO application accepted, aiming to raise ¥33 billion, with CITIC Securities and CSC as sponsors.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: The AI trade needs memory as much as compute. A ¥33 billion raise for YMTC is the market pricing storage as strategic infrastructure.

3. China’s AI4S Push: Domestic Chips Quietly Move Into Scientific Computing

   • Summary: TMTPost reports on the AI-for-science wave, where domestic accelerators increasingly power ocean forecasting and research workloads, exemplified by the full-stack domestic “Pangu” marine forecasting model.
   • Source: TMTPost · 2026-08-21
   • Editor’s Take: Science is the least forgiving benchmark for domestic chips, and it’s where they’re quietly gaining share. That is a more credible story than any spec sheet.

---

III. Humanoid Robots

1. Unitree’s Listing Frenzy: Peak ¥444.9B Market Cap on Day One

   • Summary: Unitree (宇树科技) listed on the STAR Market: IPO price ¥150.80, opened at ¥1,100 (up 629%), intraday market cap peaked around ¥444.9B, and the stock closed up roughly 460%. Founder Wang Xingxing looked visibly uneasy at the bell-ringing ceremony.
   • Source: TMTPost · 2026-08-21
   • Editor’s Take: A valuation of that size prices in years of flawless execution. The market is buying the category’s future, and the founder’s expression says he knows the scrutiny that comes with it.

2. Chaowei Power Shows the First Full Humanoid Ping-Pong Match, KAI World Model Debuts

   • Summary: At WRC 2026, Chaowei Power (超维动力) demonstrated what it calls the world’s first complete autonomous ping-pong match between humanoid robots, alongside its SMASH 2.0 system, the KAI world model, and a KAIBot body with 117 degrees of freedom.
   • Source: Leiphone · 2026-08-21
   • Editor’s Take: Table tennis is the new backflip, a demo that forces real perception latency and body control. Fun to watch, harder than it looks, and every lab knows it.

3. UBTech Shows Industrial, Commercial and Home Pipelines at WRC 2026

   • Summary: UBTech (优必选) displayed humanoid lines for factories, commercial services and home consumption at WRC 2026, promoting full-chain deployment readiness across the three scenarios.
   • Source: QbitAI · 2026-08-21
   • Editor’s Take: After clearing its embodied-model filing with the CAC, UBTech is playing the compliance-plus-lineup game. Certifications and catalogs, not demo stunts.

4. Qianxun’s WRC Exhibit: Why the “Old Demo” Criticism Misses the Point

   • Summary: QbitAI digs into Qianxun Intelligent’s (千寻智能) full-stack play at WRC, arguing the quiet, repetitive demos reveal more about real embodied capability than crowd-pleasing stunts.
   • Source: QbitAI · 2026-08-21
   • Editor’s Take: At this stage of the industry, boring reliability beats spectacular novelty. Ask how many hours the demo ran, not how high it jumped.

5. Zhongjian’s Huazhijian Debuts ZERO, a Giant Concept Humanoid

   • Summary: Zhongjian Technology’s (中坚科技) subsidiary Huazhijian (桦之坚) unveiled ZERO, a roughly two-meter-tall concept humanoid at WRC 2026 that drew the biggest crowd on the floor.
   • Source: QbitAI · 2026-08-21
   • Editor’s Take: A giant concept robot is a branding statement, not a product. Fine for buzz, but the gap between concept scale and deployment reality is exactly where the hype hides.

6. China’s Regulator Adds Embodied Robots to the “Quality Strong Chain” Push

   • Summary: China’s SAMR announced phase-two “quality strong chain” projects covering embodied-intelligence robots and solid-state batteries as strategic quality-improvement targets.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: When the quality regulator names a category, procurement standards and subsidies usually follow. Embodied robots just got put on the industrial-policy map.

7. WRC’s Trade Floor: Repair Technicians Migrate From Phones to Robot Service

   • Summary: TMTPost profiles the new livestream trade of robot maintenance, where technicians who once fixed phones and computers are retraining to service robots, with “40 days to learn” courses popping up.
   • Source: TMTPost · 2026-08-21
   • Editor’s Take: When repair classes for robots start selling, the installed base has crossed into real-world scale. After-sales is the sector’s quietest leading indicator.

---

IV. Autonomous Driving

1. Jushi + Yutong Launch Z20, a 4.2-Ton Autonomous Light Truck

   • Summary: Unmanned-delivery leader Jushi Intelligent (九识智能) and commercial-vehicle maker Yutong (宇通) unveiled the Jushi Z20 autonomous light truck with a 4.2-ton payload and 19.32 m³ cargo volume.
   • Source: QbitAI · 2026-08-21
   • Editor’s Take: Last-mile delivery moving up to 4.2-ton trucks is the tell. Autonomous freight is skipping the passenger-car detour and going straight where the margins are.

2. Qijing GX7 Debuts at Chengdu Auto Show With L3-Ready Architecture

   • Summary: The Qijing GX7 (启境 GX7) made its real-car debut at the Chengdu Auto Show with dual 15.6-inch screens, one-key-to-bed rear seats and L3-ready architecture with eight-lane redundancy.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: “L3 pre-wired” is becoming the standard spec for Chinese family EVs, a hedge that pays off the day regulations open the lane.

3. XPeng GX Passes 15,000 Deliveries in Two Months, 2,000+ Gulf Orders

   • Summary: XPeng (小鹏汽车) says the redesigned GX topped 15,000 deliveries within two months of launch, with 2,000+ orders from the UAE region, riding its second-generation VLA platform.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: Overseas order books validate the VLA-first product bet. Selling AI driving to the Gulf is the fastest revenue path China’s EV makers have found.

4. Changan Mazda EZ-60 LiDAR Edition Opens Pre-Sales

   • Summary: The Changan Mazda EZ-60 (长安马自达) LiDAR edition entered pre-sale with 1 lidar, 3 mmWave and 12 ultrasonic sensors, plus highway-and-city navigation assistance.
   • Source: ITHome · 2026-08-21
   • Editor’s Take: A joint-venture brand shipping a native lidar trim at pre-sale is the price war reaching the ADAS bin. Mid-tier autonomy is now a spec sheet, not a premium.

---

V. World Models / Physical AI

1. Xingdong’s Chen Jianyu: VLA Is Not the Endgame, Embodied Brains Enter the World-Model Era — WRC 2026

   • Summary: At WRC 2026, Xingdong Jiyuan (星动纪元) founder Chen Jianyu argued VLA models are transitional and embodied intelligence is moving to world-model-driven architectures.
   • Source: Leiphone · 2026-08-21
   • Editor’s Take: Every VLA lab is now saying the same thing: actions alone don’t generalize, world prediction does. The narrative shift is real, the architecture race is wide open.

2. Xinghaitu’s Gao Jiyang: The Endgame Business Is Selling “Physical-World Tokens” — WRC 2026

   • Summary: Xinghaitu (星海图) CEO Gao Jiyang says the ultimate business model for embodied intelligence is selling physical-world tokens, metering real-world interaction as a consumable.
   • Source: Leiphone · 2026-08-21
   • Editor’s Take: Framing robot data as tokens is a pricing model dressed as philosophy. Whatever you call it, metered physical intelligence is how this sector eventually makes money.

3. ASU’s Wei Hua: Foundation-Model Agents Are Repeating RL’s Mistakes — IJCAI 2026

   • Summary: In a Leiphone interview at IJCAI 2026, ASU professor Wei Hua (魏华) warns foundation-model agents risk repeating reinforcement learning’s history, failing to bridge the simulation-to-reality gap.
   • Source: Leiphone · 2026-08-21
   • Editor’s Take: The sim-to-real gap was RL’s graveyard, and today’s agent benchmarks have the same smell. Worth listening when a researcher says the field keeps re-buying the same lesson.

---

物界前沿 | EastVoice