EastVoice · AI · 2026-09-10 · Issue #63

Editor's Note: DeepSeek files its IPO groundwork with CITIC Securities and ships V4.1 Flash today, as Washington accuses six Chinese AI labs of "industrial-scale" distillation and Beijing's commerce ministry fires back. JD.com orders 3 million robots over five years, and Douyin just out-hours WeChat for the first time.

Commentary: The distillation fight is the most important AI story of the week, and it is not about technology. It is about who gets to define what copying a model means.

---

I. AI Foundation Models

1. DeepSeek plans to launch V4.1 Flash on September 10, claiming wins across the board over V4 Pro

• Summary: DeepSeek's open platform announced the V4.1 Flash model arrives around September 10, with internal and external tests showing improvements in performance, cost, speed, and total task time over V4 Pro. The platform also cut Flash pricing effective the same window.
• Source: ITHome · 2026-09-09
• Editor's Take: Shipping a new flagship the same week Washington names you is a statement. DeepSeek is treating release cadence as geopolitics.

2. DeepSeek hires CITIC Securities for a STAR Market IPO, filing this year

• Summary: Reuters reports DeepSeek retained CITIC Securities to prepare a Shanghai STAR Market listing, targeting an application this year and a debut next year. The company raised ¥50 billion (~$7B) in its first external round at a ¥350 billion (~$49B) valuation and is reportedly raising again.
• Source: ITHome · 2026-09-09
• Editor's Take: A company that famously took no money for two years now wants the deepest capital pool in China. The VCs who stayed out are the ones left explaining themselves.

3. Tencent's Hunyuan HY4 tops global model API call charts, up 379% week-over-week

• Summary: Tencent's Hunyuan Hy4 preview reached #1 worldwide in total LLM API calls, with weekly volume up 379%, per aggregator data cited by Geekpark. Tencent WorkBuddy simultaneously restructured the model's free-tier access ahead of commercialization on September 11.
• Source: Geekpark · 2026-09-09
• Editor's Take: Free-tier charts measure subsidized volume, not love. The real test starts when the bill arrives.

4. Ant Group open-sources Ling-3.0-flash-VL, its first native multimodal Bailing model

• Summary: Ant Group released Ling-3.0-flash-VL, a 124B-parameter MoE model activating 5.5B parameters per inference, with native image, text and video input and a 256K context window. The model introduces a visual feedback loop mechanism and is open source.
• Source: ITHome · 2026-09-09
• Editor's Take: A payments giant open-sourcing frontier-class multimodal weights keeps China's open-weights supply line running regardless of what Washington says about it.

5. MiniMax rival on-device: Facewall open-sources MiniCPM5-2B, claiming top-4B crown

• Summary: Facewall (面壁智能) and the OpenBMB community open-sourced MiniCPM5-2B, a 2B-parameter on-device model that ranks first globally among sub-4B models on the Artificial Analysis leaderboard, with tool calling, deep search and code generation tuned for phones.
• Source: ITHome · 2026-09-09
• Editor's Take: The small-model race in China is won on the phone, not the benchmark. If it runs offline on a ¥2000 handset, distribution solves itself.

6. Commerce Ministry hits back at US distillation claims: baseless, and classic double standards

• Summary: China's Ministry of Commerce responded to the joint FBI/NSA advisory accusing six Chinese AI firms of distilling US models, calling the allegations "without factual or legal basis" and an example of double standards in AI governance.
• Source: ITHome · 2026-09-09
• Editor's Take: Distillation is standard textbook technique until it happens in another country. The US just wrote the first national-security definition of a training method.

7. Huawei flagships will download and run full multimodal LLMs on-device

• Summary: Leaked specs show the upcoming Huawei Pura X View and Mate XT 2 support downloading local models of roughly 6GB (multimodal-enhanced) and 15GB (multimodal MoE) for fully on-device inference.
• Source: ITHome · 2026-09-09
• Editor's Take: Apple ships Siri features tied to a Google model. Huawei ships the model itself. In China, that difference is the whole product.

8. US-China AI traffic sits on the same table, but the money doesn't

• Summary: China's AI-native apps reached 499M monthly active users while ChatGPT holds roughly 1B weekly actives in the US, yet monetization per user diverges sharply, Tech商业 analysis finds, because the two markets sell fundamentally different products.
• Source: Huxiu · 2026-09-09
• Editor's Take: Chinese labs built world-class inference for consumers who won't pay, Americans built willingness-to-pay on top of worse latency. Both are half a business.

II. AI Software & Applications

1. Douyin's total usage time overtakes WeChat for the first time

• Summary: Nomura, citing QuestMobile, reports that in July 2026 Douyin's total user time in China surpassed WeChat for the first time. The ByteDance (字节跳动) family's share of top-50 app time rose from 33.6% to 40.9% year-over-year while Tencent's slipped to 29.1%.
• Source: Huxiu · 2026-09-09
• Editor's Take: WeChat's moat was habit, not features. A feed just ate the habit, and agents are next in line.

2. Ride-hailing meets the agent: CaoCao and Doubao launch voice-first AI taxis in three cities

• Summary: CaoCao Mobility partnered with ByteDance's Doubao assistant to ship an AI taxi service in Beijing, Hangzhou and Suzhou, where one spoken sentence sets vehicle preference, cabin temperature and destination at once.
• Source: ITHome · 2026-09-09
• Editor's Take: The app-as-portal era in China is being dismantled from the top down. Whoever owns the voice owns the order.

3. Are agents eating the app store? Chinese developers watch their entry points crumble

• Summary: A Huxiu feature argues that as users delegate tasks to agents, the mobile web's decade-old logic, where apps compete to be the user's first click, inverts: the agent decides which service gets called, not the user.
• Source: Huxiu · 2026-09-09
• Editor's Take: Every app business in China just got notified that its distribution advantage is rented, and the landlord is a model.

4. First national AI coding hackathon sets a Guinness record with 10,000 apps

• Summary: At the Bund Summit, an AI Coding hackathon drew over 10,000 participants who built 10,000+ AI applications, earning a Guinness record for most apps developed in an online generative-AI hackathon. Nearly 70% of contestants were under 18, and 11 Chinese coding platforms competed together for the first time.
• Source: Leiphone · 2026-09-09
• Editor's Take: Teenagers with copilots are the actual dataset here. China's AI-native developer pipeline starts younger than anyone assumed.

5. DeepSeek's "boyfriend mode" problem: users grief over a model upgrade

• Summary: After DeepSeek's reasoning-chain overhaul, young users who had bonded with the older persona report feeling abandoned, with some literally renaming chat windows as romantic partners, a 纵向青年 survey of user communities finds.
• Source: Huxiu · 2026-09-09
• Editor's Take: Attachment is the stickiest retention mechanic in consumer AI, and the least governed. Model upgrades are becoming breakups.

6. Bloomberg explainer puts "AI distillation" on the front burner

• Summary: Bloomberg ran a plain-English explainer on why US tech companies fear distillation, noting that a student model trained on a teacher's outputs can replicate capability at a fraction of the compute, and that detection remains difficult.
• Source: Bloomberg · 2026-09-10
• Editor's Take: When the technique is legal, cheap and nearly invisible, the only remaining moat is the one you can legislate.

III. Humanoid Robots

1. JD.com to buy 3 million robots, 1 million unmanned vehicles and 100,000 drones in five years

• Summary: At its JDD tech conference, JD Logistics announced a five-year procurement of 3M robots, 1M autonomous delivery vehicles and 100K drones, plus what it calls the world's largest robot repair-service network.
• Source: ITHome · 2026-09-09
• Editor's Take: Everyone else is pitching humanoid demos. JD just wrote the first real purchase order at logistics scale, and a service monopoly to go with it.

2. Embodied AI startups crowd past 150 as "we build robots" stops being a pitch

• Summary: A 具身志 report on China's humanoid wave notes 150+ companies now claim the space, and investors have shifted from "can you build one" to harder questions: recurring revenue, path to narrowing losses, and whether the tech is differentiated.
• Source: Huxiu · 2026-09-09
• Editor's Take: The consolidation starts before anyone turned a profit. That's how you know the capital cycle is running ahead of the product cycle.

3. Remote controls linger at China's robot games, and that's the whole critique

• Summary: Coverage of humanoid racing and weightlifting events notes nearly every competing robot still has an operator holding a teleop controller, the argument goes, until autonomy survives a doorway, embodied intelligence is a demo.
• Source: Huxiu · 2026-09-09
• Editor's Take: Sports leagues were supposed to prove the robots. Instead they're proving the pilots.

4. Chinese researchers build a living robot from bullfrog muscle

• Summary: CAS Shenyang Institute of Automation created a wireless, near-infrared-controlled manta ray robot actuated by excised bullfrog skeletal muscle, published in Advanced Functional Materials, positioning bio-actuators as a research track alongside electric motors.
• Source: ITHome · 2026-09-09
• Editor's Take: Growth will not come from muscle cells. But as long as one lab keeps exotic actuator papers flowing, the field's talent pipeline stays weird enough to matter.

5. Delivery couriers meet their steel coworkers as embodied AI absorbs ¥500 million a day

• Summary: An industry roundup counts Chinese embodied-AI funding running at roughly ¥500M (~$70M) per day in 2026, with logistics hubs the first deployment ground for humanoid helpers, even as pilot programs stall over cost and reliability doubts.
• Source: Huxiu · 2026-09-09
• Editor's Take: Warehouses are the perfect first customer: tolerable failure, measurable ROI, no press corps. The demo robots graduate here quietly.

IV. Autonomous Driving

1. Horizon Robotics and Momenta post fat gross margins, and thin profits

• Summary: H1 2026 results from Horizon Robotics and Momenta show software-style gross margins above 60-70%, but heavy R&D and per-vehicle licensing friction keep both near breakeven, framing China's AD stack suppliers as high-margin, high-effort businesses.
• Source: Huxiu · 2026-09-09
• Editor's Take: Selling intelligence per car looks like SaaS until you meet the OEMs negotiating every unit price. Margin on paper, grind in practice.

2. Aion i60 brings 192-line lidar and L4-derived end-to-end stack to mass market next week

• Summary: GAC's Aion will launch the 2027 i60 on September 15 with a 192-line lidar and a single-stage end-to-end driving algorithm the company describes as derived from its L4 program.
• Source: ITHome · 2026-09-09
• Editor's Take: Sensor premium is migrating down-market at absurd speed. Two years ago 192 lines meant a flagship; next week it means an affordable sedan.

3. Audi E7X clears 10,000 sales in 100 days, gets an "Audi world model" ADAS upgrade

• Summary: SAIC-Volkswagen says the Audi E7X became the first luxury EV to top 10,000 China sales within 100 days of launch, at 10,434 units through August, and will receive a Q4 OTA adding Audi's world-model-based assisted driving.
• Source: ITHome · 2026-09-09
• Editor's Take: The luxury badge now travels with a software roadmap. Audi is learning the Chinese way: ship first, OTA the selling points in.

4. Tesla confirms FSD V14 Lite finally coming to HW3 Model S/X

• Summary: Tesla confirmed owners of Hardware 3 Model S and Model X will receive FSD (Supervised) v14 Lite, closing a long wait after Model 3/Y HW3 cars got the feature earlier.
• Source: ITHome · 2026-09-09
• Editor's Take: Flagship owners, third in line behind the volume sellers. Compute debt from 2019 is still being repaid in release-note apologies.

V. World Models / Physical AI

1. JD launches "Physical AI Acceleration Plan" with a 100,000-card domestic cluster

• Summary: At JDD, JD Cloud president Cao Peng announced the Physical AI Acceleration Plan aimed at becoming the world's largest physical-world operations center, with JD Cloud and Moore Threads building out from a domestic 10,000-GPU cluster toward a 100,000-card AI factory.
• Source: ITHome · 2026-09-09
• Editor's Take: JD's pitch is that physical AI beats on scene access, not papers. A retailer with warehouses, drones and couriers genuinely holds the training environments nobody else can rent.

2. Wanxun's NOVA 2.0 "soft embodied brain" logs 10 million real-world jobs

• Summary: Wanxun Technology released NOVA 2.0, a neuromorphic embodied architecture it calls the first commercial-grade open-world stack for flexible manipulation, deployed across 40+ industrial scenarios with 10 million real operations logged.
• Source: Leiphone · 2026-09-09
• Editor's Take: "Ten million real jobs" is the only embodied-AI metric that means anything. Demos walk; this thing has a timesheet.

3. Ant's Lingbo robots skip the backflip, pick medicine off shelves instead

• Summary: At the Bund Summit expo, Ant Group's Lingbo Robotics drew press attention with unglamorous runs, robots picking specific medicine boxes from shelves, sustained parcel sorting, and line-loading in factories, betting the next phase of embodied AI is brains, not acrobatics.
• Source: Leiphone · 2026-09-09
• Editor's Take: The robots that don't dance are the ones getting funded. China's embodied AI press cycle has officially matured past somersaults.

4. QuestCore, NVIDIA and Intel co-author three ECCV papers on physical AI infrastructure

• Summary: Spatial intelligence firm QuestCore (群核科技) published three ECCV 2026 papers with NVIDIA, Intel and Zhejiang University building simulation and reconstruction infrastructure for physical AI, part of a batch of seven spatial-intelligence papers from Chinese labs at the conference.
• Source: Leiphone · 2026-09-09
• Editor's Take: Someone has to build the physics engines the robots dream in. It's quietly becoming a Chinese cottage industry.

---

物界前沿 | EastVoice