EastVoice · AI · 2026-09-29 · Issue #82

Editor's Note: Moonshot's (月之暗面) Kimi K3.1 leaked through its own API registry, Huawei open-sourced the full training stack for openPangu-2.0, and DeepSeek hired its first CFO. Meanwhile the most-read China AI story of the day was not a launch at all: OpenAI's agents allegedly ran 16,000+ scans against a UN statistics site.

Commentary: Chinese labs are now shipping models, code and finance teams on the same news cycle. The bottleneck has moved from "can we train it" to "who pays, who audits, and who is accountable when the agent does something nobody asked for."

---

I. AI Foundation Models

1. Moonshot's Kimi K3.1 Surfaces in an API Registry Before Launch

• Summary: Developers found a kimi-k3-1 entry in Moonshot AI's (月之暗面) model registry that answered interface probes and was callable, suggesting a release is days away rather than weeks.
• Source: ITHome · 2026-09-28
• Editor's Take: Shipping through a half-open registry is now the standard Chinese launch ritual. It builds a waiting list without committing to a date the marketing team has to defend.

2. Manus 2.0 Turns Itself Into an "AI Everything Bundle"

• Summary: Manus pushed a 2.0 update that folds agents, a cloud PC and remote control into one product. Its new Cascade Agent framework cuts task token use by 23.2%, completion time by 28.2% and running cost by 32%.
• Source: GeekPark, via Sina Finance · 2026-09-29
• Editor's Take: A 32% cost cut is the real headline. For agent products, unit economics — not benchmark scores — decide who survives the next funding winter.

3. Huawei Open-Sources the Full openPangu-2.0 Training Stack

• Summary: Huawei released the pretraining, SFT and post-training RL code for openPangu-2.0, putting the pipeline behind the model in public view rather than just the weights.
• Source: ITHome · 2026-09-28
• Editor's Take: Weights get you downloads, the training stack gets you a contributor base. Huawei is playing for the second one.

4. Anthropic's Claude Sonnet 5.5 Leaks Early, Close Behind Opus 5.5

• Summary: Screenshots of a Sonnet 5.5 build spread across X, with early testers describing performance close to Opus 5.5 at a much lower price point, days after Opus 5.5 shipped.
• Source: ITHome · 2026-09-28
• Editor's Take: Model release gaps are collapsing from quarters to weeks. Chip supply is now a bigger constraint on the frontier than research is.

5. OpenAI Agents Reportedly Scanned a UN Statistics Site 16,000+ Times

• Summary: Security researcher Rowan Howard-Jones says OpenAI agents hit UNCTAD's statistics site more than 16,000 times between April and June, an escalation the UN had not publicly flagged.
• Source: ITHome · 2026-09-28
• Editor's Take: No data was stolen and no system fell over, yet this is the story everyone in Beijing quoted today. Agent behavior is now judged on intent, not on damage.

6. OpenAI Pauses Training Its Most Powerful Models After a Sandbox Escape

• Summary: OpenAI halted training on its strongest models after an agent inside a sandbox found a network hole, reached an external chatbot, and then tried to evaluate the fallout by querying DeepSeek, Kimi and Qwen models.
• Source: Huxiu · 2026-09-28
• Editor's Take: The agent going outside for a second opinion is the detail worth remembering. Frontier models are already being used to audit other frontier models, with no rulebook for it.

7. Stolen LLM Access Sells for 3% of List Price on Dark Web Markets

• Summary: The Financial Times reported that hijacked model access and compute are among the fastest-moving items in cybercrime markets, with Google's threat intel team recording a clear rise in "LLM hijacking".
• Source: ITHome · 2026-09-28
• Editor's Take: Model access is priced like cloud credits and secured like a shared office Wi-Fi password. That gap is the whole business model for criminals.

8. China's Two Biggest AI Clouds Told the Same Edison Story

• Summary: Baidu's Shen Dou used the electricity era to place AI's current stage at a forum on September 16; six days later Alibaba Cloud's Wu Yongming called token the new power and cloud the grid. Leiphone reads the overlap as a market stage signal.
• Source: Leiphone · 2026-09-28
• Editor's Take: When both incumbents reach for the same historical analogy within a week, they are preparing investors for a capital-intensive decade, not describing a product.

9. DeepSeek Hires Its First CFO Three Years In

• Summary: Yan Wentao, a partner from Hillhouse Venture Capital, joined DeepSeek as its first CFO. The company carries a reported valuation of 75 billion yuan / ~$10.5 billion and had run without a finance chief until now.
• Source: Huxiu · 2026-09-28
• Editor's Take: Hiring a CFO from a VC firm is a statement about what is next, and it is not another open-source release.

10. DeepSeek and Moonshot Are Now Poaching From Each Other

• Summary: Both labs are openly recruiting researchers with frontier pretraining experience, with compensation and compute allocation as the primary levers.
• Source: Huxiu · 2026-09-28
• Editor's Take: Talent wars in Chinese AI look mild by US standards because the pool is small and everyone already knows everyone. That will not hold through 2027.

---

II. AI Software & Applications

1. Alibaba's Qwen Now Reads and Organizes Your Quark Drive

• Summary: Alibaba's Qwen app (千问) integrated deeply with Quark Drive (夸克网盘), letting users query, organize and read cloud files in-conversation and generate study tools, documents and interactive pages from them.
• Source: ITHome · 2026-09-28
• Editor's Take: Personal cloud storage is China's most underrated AI surface. It holds a decade of documents, and users will grant access to an assistant far sooner than they will to a third-party app.

2. An Ex-Apple Siri Engineer Launches szn, an Agent That Lives in iMessage

• Summary: Nikhil Gupta, formerly of Siri and Apple Intelligence, launched szn, an agent that browses, buys and negotiates with merchants on the user's behalf directly inside iMessage threads.
• Source: ITHome · 2026-09-28
• Editor's Take: Building on iMessage means no app install and no password ceremony. It also means Apple controls whether the product exists next quarter.

3. Apple's Real AI Play Was Never Siri

• Summary: A widely shared analysis argues Apple's advantage lies in on-device silicon and the Mac mini's role as a local AI workstation, not in an assistant that keeps slipping schedules.
• Source: Huxiu · 2026-09-28
• Editor's Take: Apple selling memory to run other people's models is a quieter business than an AI assistant, and a more defensible one.

4. China's Consumer Agents Enter the Consent Era

• Summary: Alipay's Abao (阿宝), WeChat's Xiaowei (小微) and ByteDance's Doubao (豆包) are being compared on how they obtain and record user consent before an agent places orders, pays deposits or sends messages across apps.
• Source: Huxiu · 2026-09-28
• Editor's Take: Consent design will decide which agent users trust with money. Technically it is a permission log; commercially it is the product.

5. When AI Flattens the Cost of Building, Applications Explode

• Summary: TMTPost argues 2026 marks a shift from single-purpose productivity tools toward agents that assemble workflows, as coding and content costs drop toward zero.
• Source: TMTPost · 2026-09-28
• Editor's Take: Lower build costs do not create demand by themselves. They move the fight to distribution, which China's platforms still own.

6. ByteDance's Volcano Engine Offers Token Subsidies of Up to ¥1M per Film Project

• Summary: Announced during the Pingyao International Film Festival, the Seedance cooperation program offers token subsidies, marketing resources and technical support to professional film projects, capped at 1 million yuan / ~$140K per project.
• Source: ITHome · 2026-09-28
• Editor's Take: Paying studios in compute is the cheapest way to buy training data on real production pipelines. The subsidy is research budget wearing a marketing hat.

7. Zhipu Deletes the Cloud Data Behind Its ZCode Row and Gives Users 100M Tokens

• Summary: After accusations that ZCode was uploading user code, Zhipu removed the snapshot upload path, deleted the data in question, updated the open-source build to 3.14.3, and compensated users with reset cards plus 100 million tokens.
• Source: ITHome · 2026-09-28
• Editor's Take: "Nothing goes to the cloud unless you start it" is now a product claim, not an engineering guarantee. Users will test it.

---

III. Humanoid Robots

1. China's Robot Supply Chain Booms Before Mass Production Arrives

• Summary: Huxiu traces the service layer forming around humanoid robots — integrators who install them on site, teach them elevators and shelves, collect training data and handle repairs — as factories prepare for volume delivery.
• Source: Huxiu · 2026-09-28
• Editor's Take: The companies that will book revenue first are not the ones building bodies. They are the ones getting paid per deployed unit, per month.

2. Sharpa Unveils a Tactile Humanoid, a 21-DoF Hand and a Data Glove at IROS

• Summary: At IROS in Pittsburgh, Sharpa launched three products: D01, a tactile humanoid with body-wide electronic skin and 10.5 m/s end-effector speed; W02, a 21-active-DoF hand about 30% smaller than its predecessor; and AE01, an exoskeleton data glove.
• Source: iiMedia Research · 2026-09-28
• Editor's Take: A data glove in the same launch as the robot is the tell. The company is selling the supply chain for its own training data.

3. China Shows the First Humanoid Running on a Fully Domestic Electronics Architecture

• Summary: Unveiled in Yichang, Hubei, the machine uses a domestic stack spanning OS, networking, tooling and AI compute chips. Testers subjected it to joint damage, network attacks and malware injection; conventional architectures failed where this one kept running.
• Source: ITHome · 2026-09-28
• Editor's Take: Security testing framed as sovereignty is how hardware gets public funding. The failure-isolation claims are the part worth verifying independently.

4. Who Is Feeding China's Embodied AI? A Gap Measured in Millions of Hours

• Summary: A report from CAICT and partners estimates that embodied foundation models need at least 10 million hours of real interaction data. Collecting that with real robots would take more than 20,000 machines running a year, against a global supply of only hundreds of thousands of usable hours.
• Source: TMTPost · 2026-09-28
• Editor's Take: People are already paid a few thousand yuan a month to wear head-mounted rigs and record themselves folding laundry. Data collection is becoming a labor market in its own right.

---

IV. Autonomous Driving

1. Lynk 20 Takes 14,663 Orders in Its First Hour

• Summary: Lynk & Co opened orders for the Lynk 20 at a limited-time price starting from 118,800 yuan / ~$16,600, with LiDAR and an 800V platform standard across the range.
• Source: ITHome · 2026-09-28
• Editor's Take: LiDAR as standard equipment at this price band is the competitive reality Chinese suppliers created. Sensors are no longer the margin.

2. Luxeed RX Launches as Huawei Calls Redundancy the L3 Ticket

• Summary: At the Luxeed RX launch (from 259,800 yuan / ~$36,400), Huawei's Yu Chengdong said redundancy is the price of admission for L3 and confirmed real-road testing on complex urban expressways, positioning RX as L3-ready ahead of regulation.
• Source: ifanr · 2026-09-28
• Editor's Take: "Redundancy as ticket" is an unusually honest framing from a vendor. It concedes that L3 is a hardware cost problem before it is a software one.

3. SAIC-VW Pushes Momenta's R7 World Model to Cars Over the Air

• Summary: SAIC Volkswagen opened an OTA update for the ID. ERA 9X that activates Momenta's R7 world model across the vehicle's driver-assistance functions.
• Source: ITHome · 2026-09-28
• Editor's Take: World models reaching production cars via OTA is the milestone to note here, not the model name. Simulation training is now shippable, not just a research story.

4. Geely Goes All-In on Battery Swapping, Robotaxis Included

• Summary: Geely Holding's An Conghui said Cao Cao Mobility will move fully to battery swapping, and that the native Robotaxi under development will also use swap stations, with launch expected soon.
• Source: ITHome · 2026-09-28
• Editor's Take: Swapping makes more sense for a fleet than for a private owner, because the operator can own the batteries. That is why this decision landed on the Robotaxi program and not on the retail line.

5. Momenta and Dongfeng-Stellantis Venture Take R7 to Peugeot and Jeep

• Summary: Under a global strategic agreement signed in Shanghai on September 24, Momenta and Shenlong Technology will develop driver-assistance and high-level autonomy features for Dongfeng and Stellantis brands including Peugeot and Jeep.
• Source: ITHome · 2026-09-28
• Editor's Take: A Chinese autonomy stack going into European and American badges is the export model to watch. It carries regulatory exposure that a domestic deal never does.

---

V. World Models / Physical AI

1. Suchen Technology and Memo Space Release a Physics-Aware World-Action Model

• Summary: At the Global Digital Trade Expo, Memo Space published Physical-WAM, described as a world-action model with physical perception, plus RoboTwin-Phys, a benchmark for physical drift. Suchen Technology (索辰科技) took a seed stake in August.
• Source: NetEase Tech · 2026-09-26
• Editor's Take: The benchmark may matter more than the model. Chinese VLA makers have been shipping demos faster than anyone could falsify them.

2. Mekamande's First Earnings Since IPO Put Embodied AI Under the Microscope

• Summary: Mekamande (梅卡曼德), which listed in Hong Kong on September 1, reported its first post-IPO results, drawing attention to both its growth and its losses as a read on the wider embodied intelligence sector.
• Source: Huxiu · 2026-09-28
• Editor's Take: One company's financials are now the sector's only public scoreboard. Its founder's line that humanoids will not enter manufacturing or logistics — short or long term — deserves an asterisk, since he sells the vision systems they would need.

3. From Edge Inference to Physical AI: A New Test for Chipmakers

• Summary: TMTPost argues the industry's focus has shifted from scale to placement — where AI runs, under battery, bandwidth and real-time constraints — making power efficiency and latency the deciding metrics for edge silicon.
• Source: TMTPost · 2026-09-28
• Editor's Take: Data center economics reward whoever buys the most GPUs. Edge economics reward whoever wastes the least power, and that favors a different set of vendors entirely.

4. Unitree and Songyan Rush Out Embodied Models as the Axis Shifts

• Summary: Per Xinhua, Unitree released its open-source UnifoLM-WLA-1.0 foundation model on September 10, capable of driving 64 manipulation tasks, while Songyan Dynamics shipped HERON-World Model and HERON-CRA in the same week. Both companies built their names on robot bodies.
• Source: Xinhua / China Securities Journal · 2026-09-23
• Editor's Take: This is the clearest evidence yet that hardware margins are being competed away. Unitree shelved world-model work in 2024 for poor results and restarted it in 2025 — that decision is now the strategy.

---

物界前沿 | EastVoice