EastVoice · AI · 2026-10-31 · Issue #114

Editor's Note: DeepSeek closed a second round of ¥105.5 billion (~$15.8 billion) at a ¥615.5 billion valuation, with state capital taking 27%. AgiBot shipped an embodied model it says tops three benchmarks, and Baidu open-sourced an OCR model that swallows 3,600 pages in one pass. The cheques and the demos got bigger this week. The deployment numbers did not.

Commentary: China shipped roughly 20,000 dexterous hands last year, and most of them ended up on a shelf after the demo tour. Keep that number in mind while reading the benchmark tables below.

---

I. Humanoid Robots

1. AgiBot's GO-3 Preview claims three benchmark crowns, and a shift at the convenience store

• Summary: AgiBot (智元) released GO-3 Preview on October 29, built on an embodied-native WRAM (World Reasoning Action Model) architecture that ties task reasoning, future prediction and action generation into one loop. The company says it tops three public leaderboards. Demo footage shows it making jianbing and working a convenience-store counter.
• Source: ITHome · 2026-10-30
• Editor's Take: A model that can fry an egg on camera is a benchmark result. A model that can do it 400 times before the griddle needs cleaning is a product.

2. SynDexy ships its first batch of dexterous humanoids to factory floors

• Summary: SynDexy (瞬恒智能) delivered the first small batch of its SynDexy C1 humanoid, moving from lab validation to engineering delivery. The units go into precision assembly lines at consumer-electronics and component plants.
• Source: Leiphone · 2026-10-30
• Editor's Take: Small-batch delivery is where purchase orders replace press releases. The next data point is how many units the customer reorders.

3. Tsinghua, Infinigence and AgiBot teach robots to learn from each other

• Summary: The three organisations proposed Collective Experience Learning, a training paradigm where robots share experience instead of learning only from human demonstrations. They validated it on a cluster of 96 real robots.
• Source: QbitAI · 2026-10-30
• Editor's Take: Ninety-six robots sharing one lesson is the cheapest scaling law in robotics, provided the lesson transfers across different hardware.

4. At IROS 2026, dexterous hands cleared the demo bar and hit the reliability wall

• Summary: This year's IROS floor was full of hands threading beads, inserting USB plugs, spinning pens and mixing cocktails. Leiphone's takeaway from ICRA, WRC and IROS this year is that complex motion is no longer the hard part.
• Source: Leiphone · 2026-10-30
• Editor's Take: Every one of those demos ran for ninety seconds. Production asks for eight hours, five days a week, with a maintenance log.

5. Boya's founder: 20,000 dexterous hands sold last year, most now gathering dust

• Summary: Liu Xin of Boya Intelligence (伯牙智能) estimates China shipped about 20,000 dexterous hands last year. Most went into humanoid stage shows and were then shelved. Cheap low-DOF units perform, high-DOF units break, and the expensive ones stay in the box. His view is that after GPT-6, hardware is what holds robots back.
• Source: Leiphone · 2026-10-30
• Editor's Take: An industry that sells hands for performances has a demand problem, not a dexterity problem.

6. The embodied funding squeeze is showing up in the secondary market

• Summary: TMTPost reports that AgiBot's Hong Kong listing has slowed, some holders are offering old shares at a 50% discount with no buyers, and several companies have stopped approving share reforms to keep stock from leaking out. Investors describe the pressure as concentrated after November.
• Source: TMTPost · 2026-10-30
• Editor's Take: A 50% haircut on paper is the market's way of asking which of these companies has revenue that isn't a demo contract.

7. A deployment league table for robots that already work in factories

• Summary: A survey of robots that have entered real production lines looks at cycle time, failure rates and payback periods rather than demo footage. Its finding is that one successful grasp says little about keeping pace with a line, and one installed unit says little about whether the package replicates.
• Source: Huxiu · 2026-10-30
• Editor's Take: The industry has plenty of videos and very few maintenance logs. The logs are where the market will be decided.

---

II. Autonomous Driving

1. Stelato's V8 puts Huawei's L3 architecture in a ¥329,800 sedan

• Summary: Huawei's HarmonyOS Smart Mobility lineup published options for the Stelato V8 (享界 V8). It ships with ADS 5 as standard, ±12° rear-wheel steering across the range, and an Ultra trim built on an L3 autonomous architecture. Pre-sale starts at ¥329,800, with launch expected in early November.
• Source: ITHome · 2026-10-30
• Editor's Take: L3 hardware in a ¥330,000 car is easy. The liability contract behind it is the part nobody has published.

2. Horizon's HSD stack is heading into BYD's God's Eye B

• Summary: BYD's November models will carry a new God's Eye B configuration combining Horizon Robotics' Journey 6P chip with the HSD software stack, according to Chinese media cited by ITHome. Denza, Han and high-end Ocean models are expected to follow.
• Source: ITHome · 2026-10-30
• Editor's Take: BYD just handed a domestic chip vendor its largest volume reference. That is worth more than any benchmark score.

3. Zhuoyu takes its driving stack to trucks, and promises a one-year payback

• Summary: Zhuoyu Technology (卓驭) launched Panxing, a commercial-vehicle assistance system split into PDS 4 for highway NOA on heavy trucks and PDS 5 for buses, light trucks and robovans. The company claims operators can recover the cost within a year.
• Source: Leiphone · 2026-10-30
• Editor's Take: Trucking fleets buy on fuel and insurance math, not on autonomy narratives. A one-year payback claim will be audited by CFOs, not analysts.

---

III. World Models & Physical AI

1. IM's LS8 puts a world model on a ¥250,000 SUV

• Summary: SAIC's IM (智己) launched the LS8 from ¥250,000, with a steer-by-wire chassis and CATL batteries standard across the range. Momenta's world model runs on the car, part of a pitch that includes lifetime AI features.
• Source: QbitAI · 2026-10-30
• Editor's Take: World models are now a trim-level feature. The interesting question is what happens to the car's behaviour when the model is updated six months after purchase.

2. Li Fei-Fei's three pits, and the $8.2 billion anchor under the world-model trade

• Summary: A Leiphone profile traces Fei-Fei Li's path from ImageNet to spatial intelligence and the three gaps she has to close. It notes that $8.2 billion in capital has already set a valuation anchor for the world-model sector.
• Source: Leiphone · 2026-10-30
• Editor's Take: ImageNet worked because the labels were cheap and the task was narrow. Spatial intelligence has neither advantage.

3. A Hangzhou data centre wants to compress two years of embodied data into one month

• Summary: Yingsu (影溯) opened what it calls China's first embodied simulation data-collection centre, in Hangzhou. The industry needed roughly two years to accumulate a million hours of robot data. Yingsu says it can produce that in a month.
• Source: Leiphone · 2026-10-30
• Editor's Take: Synthetic hours are cheap. The bill arrives when a policy trained in simulation meets a greasy kitchen counter.

---

IV. Foundation Models

1. DeepSeek raises ¥105.5 billion at a ¥615.5 billion valuation

• Summary: DeepSeek's second round brought in ¥105.5 billion (~$15.8 billion) from about 30 investors, according to reports cited by TMTPost. State capital accounts for 27%, VCs 26.5% and tech companies 41.9%, including Tencent, CATL, CXMT, SMIC, Meituan and vivo. It is the largest venture round ever raised by a Chinese AI company.
• Source: TMTPost · 2026-10-30
• Editor's Take: A cap table that mixes chip fabs, battery makers and a video platform is an industrial policy instrument, not a venture round.

2. SiliconFlow files for a Hong Kong listing

• Summary: Token supplier SiliconFlow (硅基流动) has had its Hong Kong IPO filing accepted by the CSRC. The company reports over 10 million registered users and 13,000 enterprise customers, and raised close to ¥2.9 billion this year. Frost & Sullivan ranks it as China's largest independent token supplier.
• Source: ITHome · 2026-10-30
• Editor's Take: Routing tokens is a thin-margin business until you own the capacity underneath. The filing will show which side of that line SiliconFlow sits on.

3. DeepSeek's front-end code leaks a cybersecurity model called Cyber

• Summary: Developers found application routes and a feature gate for a Cyber model in DeepSeek's site code, pointing to enterprise and personal editions in limited grey release. DeepSeek has not announced the product.
• Source: ITHome · 2026-10-30
• Editor's Take: A security model is the one product where a leak before launch is itself a security story.

4. Baidu open-sources an OCR model that reads 3,600 pages in one pass

• Summary: Baidu released ERNIE Unlimited OCR 2 on October 27, removing the page ceiling that constrains document parsers. In its demo it processed all seven Harry Potter books, more than 3,600 pages, in a single pass.
• Source: ITHome · 2026-10-30
• Editor's Take: Long-context OCR is unglamorous and immediately useful. That combination is rare in Chinese model releases this year.

5. Xiaomi open-sources a 0.8B OCR model for documents and visual Q&A

• Summary: Xiaomi released a 0.8B parameter OCR model that handles document parsing and visual question answering in one model, with Markdown and JSON output.
• Source: ITHome · 2026-10-30
• Editor's Take: Small models on-device are how phone makers avoid paying per-token rent to a cloud they do not control.

6. Iluvatar's Tiangai 300 matches H200 in hands-on tests

• Summary: At a media day this week, Iluvatar CoreX (天数智芯) benchmarked its Tiangai 300 against Nvidia's H200 and reported comparable performance. The company is now pitching heterogeneous systems and an agent-based software stack rather than raw FLOPS.
• Source: Leiphone · 2026-10-30
• Editor's Take: Matching H200 on one workload is a milestone. Matching CUDA's decade of tooling is the actual product.

7. Qwenputer puts a 32-bit RISC-V CPU inside a language model

• Summary: Developer Divyansh Agrawal released Qwenputer on October 22, built on Qwen3.5-4B. The model emulates a 32-bit RISC-V CPU and memory inside its own network, so it can execute programs without calling Python or any external tool.
• Source: ITHome · 2026-10-30
• Editor's Take: Teaching a model to be its own interpreter is a neat trick. Whether it computes arithmetic more reliably than a calculator is a separate question.

---

V. AI Software

1. Apple Intelligence is finally clearing Chinese regulatory ground

• Summary: Apple said on October 31 that the first batch of Apple Intelligence features aimed at improving the iPhone experience is coming to users in China, without giving a full feature list or a date for every capability.
• Source: ITHome · 2026-10-31
• Editor's Take: Two years after launch, Apple's China AI story is still a schedule announcement. Domestic rivals shipped on-device assistants while it waited.

2. Huawei's XiaoYi gets memory and triggers, with 100+ apps wired in

• Summary: Huawei upgraded XiaoYi on the Mate 90 series, with two new capabilities: XiaoYi Bangji (帮记), which remembers context across apps, and trigger-based AI that fires on events. More than 100 apps have integrated the APIs.
• Source: ITHome · 2026-10-30
• Editor's Take: Trigger-based agents are the pragmatic version of the assistant dream. The phone acts without being asked, which is also the version users will complain about.

3. Baidu rebuilds search around recursive self-improvement, and cuts API prices

• Summary: Baidu announced four AI upgrades to its app on October 30, including a search stack it describes as recursively self-evolving, a rebuilt search API priced at ¥4.9 per 1,000 calls during the campaign, and new camera and GUI features. That undercuts DeepSeek's ¥5 standard rate.
• Source: QbitAI · 2026-10-30
• Editor's Take: Baidu is pricing search calls below a competitor that does not sell search. That tells you where the margin is going.

4. Xiaohongshu buys travel-planning app Pi Travel outright

• Summary: Registration records show Xiaohongshu (小红书) has become the sole shareholder of the company behind Pi Travel (圆周旅迹), an AI itinerary app, with founder Zheng Yifan exiting. Registered capital is ¥1 million.
• Source: TMTPost · 2026-10-30
• Editor's Take: A travel planner is the shortest path from a saved post to a booked trip. Xiaohongshu bought the last mile.

5. Shagang puts Qwen agents into electric-furnace and finance workflows

• Summary: Jiangsu Shagang Steel has deployed Qwen Office agents in production, covering electric-furnace review, transformer condition assessment and financial analysis. The company says some core scenarios are now hundreds of times faster.
• Source: Leiphone · 2026-10-30
• Editor's Take: Steel mills are a better AI market than law firms. The workflows are documented, the data is instrumented, and nobody argues about creative license.

6. Alibaba Cloud's Qoder passes $100 million in recurring revenue

• Summary: Qoder, Alibaba Cloud's AI coding product, has crossed $100 million in annualised recurring revenue, and the team is splitting the product line. The number was first discussed at the September Yunqi conference.
• Source: TMTPost · 2026-10-30
• Editor's Take: Coding assistants are the one Chinese AI category with paying enterprise customers at scale. Qoder is now the number to beat.

7. Two Chinese AI companion apps vanish from the US App Store

• Summary: In mid-October, PolyBuzz, owned by Zuoyebang, was pulled from the US and UK App Stores on October 15, and a second Chinese AI companion app disappeared from the US store at the same time. The products were reportedly earning over ¥100 million a month.
• Source: Huxiu · 2026-10-30
• Editor's Take: Companion apps were the fastest Chinese AI export to make real money abroad. Apple just showed how quickly that channel can be closed.

---

物界前沿 | EastVoice