EastVoice · AI · 2026-08-15 · Issue #37

Editor’s Note: Alibaba open-sourced Qwen3.8-27B, a mid-size model that runs on consumer GPUs while beating Qwen3.7-Plus on coding and office tasks, as total Qwen downloads pass 3 billion. Zhipu shipped GLM-5.3, claiming coding gains of roughly 50% over GLM-5.2 on an unchanged base. And a former Meta AI researcher left a 7-digit USD salary to start something new.

Commentary: Chinese labs are widening the “open but cheap” front: every week a new open-weight release lands with consumer-hardware hooks. The downloads race is one thing, converting that into enterprise revenue is the next stage.

---

I. AI Foundation Models

1. Alibaba open-sources Qwen3.8-27B: runs on consumer GPUs, beats Qwen3.7-Plus on coding and office

   • Summary: On August 14, 阿里云 / Alibaba Cloud released Qwen3.8-27B as open weights. The 27B-class model is designed to fit on consumer-grade GPUs while outperforming the larger Qwen3.7-Plus on programming and office benchmarks, per ITHome and QbitAI. Total Qwen downloads across the ecosystem passed 3 billion.
   • Source: ITHome · 2026-08-14
   • Editor’s Take: The “runs on my home GPU” pitch is the sharpest weapon Chinese open-weight labs have. It turns every developer with a gaming rig into an Alibaba ambassador.

2. Zhipu ships GLM-5.3: unchanged base, ~50% coding gain, and it surfaced a 40-year-old bug

   • Summary: 智谱 / Zhipu released GLM-5.3, claiming coding performance roughly 50% better than GLM-5.2 on the same base architecture, inching closer to Anthropic’s Claude Fable 5. The demo highlight: the model reportedly found a 40-year-old security bug during testing.
   • Source: Leiphone · 2026-08-14
   • Editor’s Take: “Same base, better model” is efficiency engineering, not a new frontier — but a 50% coding jump without a new architecture is exactly the kind of moat China’s labs are building quietly.

3. Qwen download total crosses 3 billion, per Alibaba’s open-source push

   • Summary: 雷锋网 / Leiphone reports Alibaba’s 千问 / Qwen family has now accumulated over 3 billion downloads globally, riding a string of open-weight releases this year and expanding use in enterprise and academic deployments.
   • Source: Leiphone · 2026-08-14
   • Editor’s Take: Downloads are a vanity metric until they become revenue, but 3 billion is the sort of scale Western labs can no longer ignore when pricing their own APIs.

4. Yujia Yu leaves Meta after a reported ¥700M-equivalent package to start a venture

   • Summary: 量子位 / QbitAI reports AI researcher 余家辉 / Yujia Yu departed Meta — where he reportedly earned a package worth around 700 million RMB — to found his own startup. The story is being read as a premium on scarce Chinese AI talent.
   • Source: QbitAI · 2026-08-14
   • Editor’s Take: When money can’t hold talent, it usually means the person believes their equity in a new thing beats someone else’s equity in a giant. Watch where this one lands.

5. China’s National Standards kick off humanoid-robot testing framework, Unitree and Xiaomi attend

   • Summary: The working group for national standards on humanoid robot test methods held its kickoff in Wuhan, with Unitree (宇树科技) and Xiaomi Robotics among participants, per ITHome. The standard aims to unify evaluation across safety, motion, and interaction.
   • Source: ITHome · 2026-08-14
   • Editor’s Take: Standards work is the boring half of robotics industrialization, but it’s how you get from lab stunts to insurable, deployable machines.

II. AI Software & Applications

6. Guangdong issues its first “Token Loan”: credit line of ¥28M in pilot

   • Summary: 广东 / Guangdong unveiled its first “Token Loan” — a financing product where AI computing tokens serve as the collateral basis — with pilot credit lines reaching 28 million RMB, per ITHome. It’s an attempt to monetize AI usage itself as an asset class.
   • Source: ITHome · 2026-08-14
   • Editor’s Take: Lending against tokens is a creative way to bankroll startups that are burning compute, not real estate. Whether banks can price token value is another question.

7. MiniMax ships Music3: AI-generated songs up to 5 minutes

   • Summary: 稀宇科技 / MiniMax released MiniMax-Music3, an audio model capable of generating songs up to 5 minutes long in a single pass, covering vocals and instrumentation, per ITHome.
   • Source: ITHome · 2026-08-14
   • Editor’s Take: Music generation is becoming a consumer product, not a demo. Five minutes is long enough for real song structure — the licensing questions will follow fast.

8. Baidu renames GenFlow to “Kuku AI”, ships standalone office app

   • Summary: 百度 / Baidu’s 文库网盘 / Wenku-Pan GenFlow rebranded as 「库库AI」(Kuku AI), now claiming 25 million+ monthly active users for its AI office suite, and released a standalone desktop client, per Leiphone and QbitAI.
   • Source: Leiphone · 2026-08-14
   • Editor’s Take: Baidu keeps trying to ship AI productivity beyond search. 25M MAU is real traction — the question is whether it converts to paid seats.

9. Moonshot AI issues stern statement over suspicious trading activity

   • Summary: 月之暗面 / Moonshot AI published a rare public statement warning of suspicious transactions targeting the company and said it has reported the matter to public security authorities, per ITHome.
   • Source: ITHome · 2026-08-14
   • Editor’s Take: A public complaint to police is usually the last resort after private channels failed. Read it as a signal that China’s AI funding frenzy has attracted unwanted actors.

III. Humanoid Robots

10. Robots’ compute hunger: demand seen rising 10x in two years

    • Summary: 量子位 / QbitAI reports that compute demand for embodied AI has grown roughly 10x in two years, as humanoids move from scripted demos to real-world perception and planning.
    • Source: QbitAI · 2026-08-14
    • Editor’s Take: Everyone prices robot hardware; almost no one prices the data center behind it. The 10x figure is the hidden business in the robot story.

11. Honor teases a “Robot Phone” beyond current form factors

    • Summary: 荣耀 / Honor’s imaging chief 罗巍 / Luo Wei told ITHome that next year the company will ship “a product nobody has seen or even imagined,” following its earlier Robot Phone first-gen exploration. Honor’s other executive 方飞 / Fang Fei said phones will move from “containers of apps” to “stages for agents.”
    • Source: ITHome · 2026-08-15
    • Editor’s Take: Chinese vendors love a bold phone bet. “Agent-native phone” is a good story; shipping one that replaces the app grid is the hard part.

12. Geely’s Eva in-car robot hits JD.com at ¥799

    • Summary: 极氪 / Zeekr’s Eva in-car companion robot went on sale on JD.com with an early-bird price of 799 RMB, featuring active tracking and directional steering, per ITHome.
    • Source: ITHome · 2026-08-14
    • Editor’s Take: In-car robots at sub-¥800 price points are testing whether consumers actually want a dashboard pet. Cheap enough to try, easy enough to ignore.

IV. Autonomous Driving

13. Pony.ai x Uber: 2,000 robotaxis planned across 5 European cities

    • Summary: 小马智行 / Pony.ai expanded its partnership with Uber to deploy more than 2,000 robotaxis across five European cities, pairing Pony’s autonomous stack with Uber’s ride-hailing network, per ITHome.
    • Source: ITHome · 2026-08-14
    • Editor’s Take: Europe is the prize China’s AV players have struggled to reach. Uber gives Pony distribution overnight; Pony gives Uber a moat against Waymo. Neat trade.

14. WeRide H1 revenue ¥350M, +73% YoY, 1,800+ robotaxis

    • Summary: 文远知行 / WeRide reported H1 2026 revenue of 350 million RMB, up 73% year-over-year, with its robotaxi fleet exceeding 1,800 vehicles, per ITHome.
    • Source: ITHome · 2026-08-14
    • Editor’s Take: 1,800 robotaxis is real operational scale, and 73% growth is a healthy number for a listed AV company in a sector full of promises.

15. Chipone’s automotive 7nm AI accelerator “Tiangong 100” enters mass production

    • Summary: 芯擎科技 / Chipone Technology said its automotive-grade, 7nm AI accelerator 「天工 100」(Tiangong 100) has entered mass production, targeting in-vehicle AI workloads at 千元级 / ~¥1,000-class hardware cost, per ITHome.
    • Source: ITHome · 2026-08-14
    • Editor’s Take: Cheap automotive AI silicon is the supply-chain story of Chinese EVs. At ¥1,000-class cost, every car becomes an AI device by default.

V. World Models / Physical AI

16. Aibee’s Lin Yuanqing: “build strong forts, fight steady battles” toward Physical-AGI

    • Summary: 爱笔智能 / Aibee founder 林元庆 / Lin Yuanqing told 钛媒体 / TMTPost that the company is pursuing物理世界AGI (physical-world AGI) with an “undramatic” playbook: dense deployments in retail and airports first, then generalization.
    • Source: TMTPost · 2026-08-14
    • Editor’s Take: Aibee’s “boring first” strategy is the opposite of the demo-driven hype cycle. It’s the honest way to build physical AI, and the slowest to fund.

17. Qianhe Yibang closes ¥2B+ Series B for 3D compute-in-memory chips

    • Summary: 谦合益邦 / Qianhe Yibang, a 3D compute-in-memory (存算一体) chip maker, completed a Series B of over 2 billion RMB, per Leiphone, to scale its monolithic 3D CIM products for AI inference.
    • Source: Leiphone · 2026-08-14
    • Editor’s Take: Compute-in-memory is Silicon Valley’s lab curiosity and China’s funding pattern. Big money is betting the memory wall makes CIM inevitable, eventually.

---

物界前沿 | EastVoice