EastVoice · AI · 2026-10-28 · Issue #112

Editor's Note: Xiaomi showed a 21-DoF bionic hand built on 1,827 real factory procedures, Unitree spent the week defending a ¥271.8 billion market cap, and Baidu open-sourced an OCR model that holds its KV cache constant no matter how many pages you feed it. On the money side, HAI Robotics filed for a Hong Kong listing carrying ¥5.2 billion in accumulated losses, and the Hurun rich list put two AI founders at the top for the first time.

Commentary: China's embodied-AI trade has moved from "can it demo" to "can it invoice." The companies now under pressure are the ones that booked orders before they had a product to ship.

---

I. Humanoid Robots

1. Xiaomi's new hand has 21 degrees of freedom and 1,500 touch points

• Summary: Xiaomi's manufacturing arm disclosed a bionic hand aimed at factory work: 21 degrees of freedom, 26 motors under hybrid drive, and a 1,500-point tactile array across the palm. The design was tuned against 1,827 real standard operating procedures collected at the pilot center inside its auto plant, which is the part that matters. Most dexterous hands are built for a lab bench; this one was built against a line schedule.
• Source: ITHome · 2026-10-27
• Editor's Take: The interesting number here is not 21. It is 1,827. A hand that has seen eighteen hundred real SOPs has a spec sheet written by production engineers, not by a marketing team.

2. Unitree gains 7.69% in a day and still has to answer for its valuation

• Summary: Unitree closed at ¥672 a share on October 26, a one-day gain of 7.69%, putting its market cap back near ¥271.8 billion. The rebound revived the argument that has followed the company since listing: what, exactly, is being priced here. Revenue, order book, or the option value of an entire industry.
• Source: Huxiu · 2026-10-27
• Editor's Take: A humanoid maker trading at a multiple that assumes mass adoption is a bet on 2029 placed in 2026. Retail investors are effectively buying a call option and calling it a stock.

3. One humanoid startup just open-sourced everything, down to the algorithms

• Summary: A Chinese humanoid company released its full stack, mechanical drawings included, from joint design through the locomotion control code. Download and build is the pitch. Full-stack open sourcing is rare in this category for an obvious reason: the hardware drawings are the only moat most of these teams have.
• Source: QbitAI · 2026-10-27
• Editor's Take: Open-sourcing your hardware is a confession that your edge is not in the hardware. Sometimes that is honest, sometimes it is a fundraising tactic aimed at developers rather than buyers.

4. HAI Robotics heads for Hong Kong with ¥5.2 billion of accumulated losses

• Summary: Shenzhen-based HAI Robotics filed for a Hong Kong IPO. Revenue grew at a 58% compound annual rate, and the company says it held more than 30% of the global warehouse-robotics market in 2025 by both revenue and units shipped. Cumulative losses since founding reached ¥5.23 billion as of June 2026, with ¥3.52 billion of that booked in the last three and a half years.
• Source: TMTPost · 2026-10-27
• Editor's Take: Thirty percent of the global market and still burning cash is the warehouse automation story in one line. The buyers are logistics giants who negotiate like logistics giants, and the hardware gets cheaper every year.

5. A humanoid learns shelf-stocking by failing at it, with nobody in the loop

• Summary: A robot working a real convenience store figured out how to hang merchandise on a hook that was partly hidden behind a price tag. It missed on the first attempt, then corrected without a human operator stepping in. That loop, deploy, fail, retry, keep the weights, is what physical reinforcement learning looks like outside a simulation.
• Source: QbitAI · 2026-10-27
• Editor's Take: Long-tail failures are where every retail deployment dies. Solving them in a live store, without a teleoperator babysitting, is a more meaningful milestone than any walking demo.

6. Samsung's chairman tours the robotics division before CES 2027

• Summary: Lee Jae-yong made his first visit to Samsung's RX robotics division, reviewing technology headed for CES 2027. Samsung has been slower than Hyundai and LG to put a humanoid story in front of investors, and this visit reads as a scheduling decision rather than a product one.
• Source: ITHome · 2026-10-27
• Editor's Take: A chairman visit is a budget signal. Expect the actual hardware in January, and expect it to be positioned as a platform play rather than a robot.

---

II. Autonomous Driving

1. WeRide signs Allianz and Ping An for a global insurance program

• Summary: WeRide launched a global insurance partnership covering its existing operations in China, Southeast Asia and the Middle East, with Allianz Insurance Singapore and Ping An Property & Casualty as the first two carriers. The program is designed to extend to Europe as permits come through.
• Source: ITHome · 2026-10-27
• Editor's Take: Insurance is the least glamorous gate in autonomy and the one that decides the unit economics. A carrier willing to underwrite you in three regions is worth more than a press release about mileage.

2. Deepal S09 gets Huawei's ADS Max V5.0.2 over the air

• Summary: Changan's Deepal pushed OTA V5.4 to the S09, the first model to carry Huawei's Qiankun ADS Max V5.0.2. The update bundles four categories of change and more than 20 new features.
• Source: ITHome · 2026-10-27
• Editor's Take: Huawei's stack now ships as a versioned software product with a release cadence. That is a bigger shift in the supplier landscape than any single feature in the changelog.

3. Leapmotor pushes park-to-park driving to its lidar trims

• Summary: Leapmotor began rolling out an OTA for 2026 C-series models, adding park-to-park assisted driving on lidar-equipped trims. Covered models include the C10, C11 and C16, with a handful of entry variants excluded.
• Source: ITHome · 2026-10-27
• Editor's Take: Park-to-park is becoming table stakes in the ¥150,000 bracket. The differentiator is no longer whether the feature exists but how often the driver has to take over mid-route.

4. Neolix takes strategic money from Xinhua Holdings

• Summary: Neolix signed a strategic investment agreement with Xinhua Holdings, the investment arm of Xinhua News Agency, marking that group's first direct industrial bet on autonomous driving and physical AI. Neolix says it became the industry's first ¥10 billion unicorn and the first to reach volume production of a mapless L4 stack.
• Source: Leiphone · 2026-10-27
• Editor's Take: State-affiliated capital entering delivery robots is a signal about procurement, not technology. The money is less interesting than the logistics contracts it usually arrives with.

5. ADAS suppliers are shipping more units and keeping less of the money

• Summary: As assisted driving becomes standard equipment, the competitive bar for suppliers has shifted from algorithm demos to delivery reliability, chip portability and cost control. TMTPost's read is that winning a nomination now depends on hitting a price target and a launch date across multiple vehicle platforms, not on having the best feature list.
• Source: TMTPost · 2026-10-27
• Editor's Take: This is what a market looks like after the technology stops being scarce. The survivors will be the ones who treated automotive qualification as their core product rather than as paperwork.

---

III. World Models & Physical AI

1. Ant Group's embodied-model contest drew 1,855 teams from 13 countries

• Summary: The online round of the first Ant Lingbo Embodied Model Challenge closed with 1,855 teams and 3,252 contestants registered, drawn from 648 universities, research institutes and companies across 13 countries, including 41 overseas institutions. Finals are next.
• Source: Leiphone · 2026-10-27
• Editor's Take: Contests are how a field without a benchmark builds one. The real output is not the prize winner but the shared evaluation harness everyone keeps using afterwards.

2. The ¥1 billion order book that stopped a company

• Summary: Nanjing Wujing Future Power, founded in February 2026, was reported to be at a standstill after claiming nearly ¥1 billion in orders. ¥500 million of that was a framework agreement with a central state-owned enterprise and another ¥500 million was local market intent, neither of which is a purchase order.
• Source: Huxiu · 2026-10-27
• Editor's Take: Framework agreements are the most overvalued asset class in Chinese hard tech. A letter of intent from a state-owned buyer has brought down more startups than any technical failure.

3. Physical AI moves onto the factory floor, one changeover at a time

• Summary: At the 26th China International Industry Fair in Shanghai, "factory plus physical AI" was the dominant theme. Han Chen, president of ABB Robotics China, argued that high-mix, low-volume production with frequent line changeovers is where the technology actually pays off, since that is the case fixed automation handles worst.
• Source: Huxiu · 2026-10-27
• Editor's Take: Every credible physical-AI pitch in China now starts with small batches and frequent changeovers. It is the one workflow where a learning system beats a hard-coded one on day one.

---

IV. Foundation Models

1. iFlytek trains X2.6 entirely on domestic silicon and bets on the edge

• Summary: iFlytek released Spark X2.6 as a cloud-plus-device pair, a large cloud model alongside a 4B on-device model that takes tasks the cloud does not need. Chairman Liu Qingfeng said domestic compute still trails Nvidia by roughly two generations, and predicted parity within two years.
• Source: TMTPost · 2026-10-27
• Editor's Take: Two generations behind is an admission, and a strategic one. If export controls hold, the on-device path is where a Chinese vendor can compete without waiting for the fab problem to be solved.

2. GPT-6.1 Terra at one-tenth the price, with Astra-class computer use

• Summary: Early testers ran OpenAI's GPT-6.1 Terra through desktop control tasks and came back reporting Astra-level computer use at roughly 10% of the previous price point. OpenAI shipped it the same day it published three scientific results, which is not a coincidence in a week when Anthropic was filing for an IPO.
• Source: QbitAI · 2026-10-27
• Editor's Take: Price cuts on agentic models are a land grab for the workflow layer. Whoever owns the desktop session owns the switching cost, and that is worth more than the inference margin.

3. ByteDance has the most compute and the slowest models, and the org chart explains it

• Summary: TMTPost traces ByteDance's model gap to its internal horse-racing structure, where multiple teams compete on overlapping goals. The mechanism that produced Douyin out of a crowded app market works differently when every experiment needs the same scarce GPUs and the same data pipeline.
• Source: TMTPost · 2026-10-27
• Editor's Take: Internal competition optimizes for shipping fast and cheap. Frontier model training optimizes for not restarting a run that costs eight figures, and those two instincts are not compatible.

4. Baidu open-sources an OCR model with a constant KV cache

• Summary: Baidu released Unlimited OCR 2, an open-source document parser that keeps its KV cache roughly constant as page count grows, cutting cache usage by 96% and removing the practical page limit. The demo run fed it seven volumes of Harry Potter in one pass.
• Source: QbitAI · 2026-10-27
• Editor's Take: Long-document parsing was a memory problem dressed up as a model problem. Fixing the cache is unglamorous engineering, and it is what makes enterprise document workflows actually deployable.

5. Hi3D ships an open 3D model that is cleared for commercial use

• Summary: A Chinese team released Hi3D, a 3D generation model with weights available and a commercial license. The pitch is narrow specialization: a 3D-native model beating general multimodal models at their own 3D tasks.
• Source: QbitAI · 2026-10-27
• Editor's Take: General models keep getting better at everything except the things where the data pipeline is the product. 3D assets are one of those, for now.

6. MiniMax's cheap model meets an awkward question about memorized benchmarks

• Summary: After MiniMax released M3.1 Flash on October 19, attention moved from its leaderboard position to how it got there. One widely shared evaluation found the model near-perfect on older test items and near-zero on new ones, the signature of a model that has seen the answers.
• Source: TMTPost · 2026-10-27
• Editor's Take: Contamination is now the default explanation for a surprising score, which means every benchmark number published in China this quarter deserves a second read.

7. The leaderboard business is being repriced

• Summary: TMTPost argues that AI rankings are entering a credibility reset, as buyers shift from "what did it score" to "what did it deliver in my workflow." Vendors that built their launch around a first-place ribbon are the ones most exposed.
• Source: TMTPost · 2026-10-27
• Editor's Take: Benchmarks were always marketing with a methodology section. What changed is that the buyers now have their own evals, and they are not publishing them.

---

V. AI Software

1. China's new rich list is an AI list

• Summary: The 2026 Hurun Rich List, with wealth measured as of September 1, puts ByteDance founder Zhang Yiming at ¥670 billion and DeepSeek's Liang Wenfeng at ¥280 billion. It is the first edition where AI fortunes, rather than property or consumer internet, set the top of the table.
• Source: Huxiu · 2026-10-27
• Editor's Take: Paper wealth in a private AI lab is a statement about expected exits, not realized value. DeepSeek in particular has no funding round to mark against, so treat that ¥280 billion as a guess with a byline.

2. Ant Group's Lingguang 2.0 wants to manage your errands, not answer them

• Summary: Ant Group released Lingguang 2.0, repositioning the assistant from single-turn responses to ongoing task management. The pitch is an agent that notices what needs doing, tracks it across sessions and closes it out, rather than one that waits for a prompt.
• Source: ITHome · 2026-10-27
• Editor's Take: Every assistant in China now claims persistent memory. The differentiator will be what happens when it remembers something wrong and acts on it anyway.

3. Manus marks its second version with a manifesto rather than a feature list

• Summary: On October 26, Manus co-founder and product partner Zhang Tao laid out the company's 2.0 direction. Notably absent was a model upgrade. The framing was about how work gets delegated and reviewed, which is where agent products keep losing users after the first week.
• Source: GeekPark · 2026-10-27
• Editor's Take: A manifesto is what a company writes when the technology has plateaued and the workflow has not. That is a reasonable place to compete from.

4. Qwen Office runs on Rokid's glasses

• Summary: At its launch event on October 27, Rokid announced that Alibaba's Qwen Office is now available inside its AI glasses. Users add the agent in the Rokid AI app, link their Qwen Office account, and issue tasks by voice with results surfaced on the lens.
• Source: Leiphone · 2026-10-27
• Editor's Take: Glasses are a bad place to read documents and a good place to start a task. Treating the lens as a command surface rather than a screen is the only version of this that works.

5. Haier wins the country's first national AI pilot base for home appliances

• Summary: Haier began construction on a national AI application pilot base for home and kitchen appliances in Qingdao, recommended by the Ministry of Commerce and approved by the NDRC, with more than 200 industry experts attending. The base is meant to push cross-brand interoperability, which is the part appliance makers have historically resisted.
• Source: ITHome · 2026-10-27
• Editor's Take: A state-backed pilot for cross-brand interoperability is an admission that the market failed to produce it. Haier hosting the effort is either a concession or a very good position to shape the standard.

6. Robotics pay in China: ¥100 million for a chief scientist, ¥100,000 a month for interns

• Summary: Embodied AI companies are bidding up talent as rounds close and lines start up. TMTPost reports top scientists being offered packages around ¥100 million, while some firms pay interns and PhD candidates ¥100,000 a month to keep the pipeline warm.
• Source: TMTPost · 2026-10-27
• Editor's Take: Salary inflation this early in a hardware cycle usually means too much capital chasing too few people who have shipped a working system. It also means the burn rate projections in those IPO filings deserve scrutiny.

---

物界前沿 | EastVoice