EastVoice · AI · 2026-10-29 · Issue #112

Editor's Note: Unitree posted its first results since listing, and the shape of the quarter is the story: revenue up nearly 70%, profit down, R&D up about 2.5x. XPeng says solid-state cells for its IRON humanoid are coming off the line in small batches. Volkswagen put a 2027 date on L3 in China and confirmed it is designing its own L4-grade SoC.

Commentary: China's robot trade has spent two years pricing the future. This week the invoices arrived, and they belong mostly to carmakers and parts suppliers.

---

I. Humanoid Robots

1. Unitree's first post-IPO quarter: revenue up 70%, profit going the other way

• Summary: Unitree reported nine-month revenue of ¥1.817 billion, up 55.71%, and net profit of ¥387 million, up 270.54%. The third quarter is where the line bends. Revenue grew nearly 70% year on year while quarterly net profit fell, and R&D spending rose roughly 2.5x. The nine-month figures look strong because the first half carried them.
• Source: ITHome · 2026-10-28
• Editor's Take: A hardware company that doubles R&D while quarterly profit slips is still buying its way into a product cycle. Investors will forgive that for two more quarters. After that, the R&D has to show up as units shipped.

2. XPeng has solid-state cells for its IRON robot coming off the line

• Summary: Group vice president Gu Jie, speaking at the SAECCE conference, said the solid-state battery for XPeng's IRON humanoid has entered small-volume production and the robot is now in the stage before SOP. Solid-state buys runtime per charge and frees torso volume that a conventional pack eats.
• Source: ITHome · 2026-10-28
• Editor's Take: The notable part is not the chemistry. It is that a carmaker is running the cell line, which means the yield data will be audited to automotive standards rather than to a demo schedule.

3. Unitree and Sequoia China start a joint vehicle for embodied AI startups

• Summary: The two announced the "Shanyu Venture Program", aimed at seed and early rounds across embodied AI, robot applications and upstream components. Unitree frames its role as a co-founder rather than a passive cheque.
• Source: ITHome · 2026-10-28
• Editor's Take: When a hardware maker starts funding its own supply chain, it is buying priority on parts before they get scarce. It also means Unitree's balance sheet now has a stake in deciding which competitors survive.

4. A Sichuan convenience chain tried robot clerks. The robots lost.

• Summary: One of Sichuan's largest community convenience chains put humanoid robots behind the counter in early 2025 to sell drinks and snacks. QbitAI's account of the wind-down is a list of ordinary failures: slow grasping, aisles blocked by a machine that would not yield, shelves no demo had prepared it for.
• Source: QbitAI · 2026-10-28
• Editor's Take: Retail was billed as the easy on-ramp, with fixed shelves, repeatable SKUs and hours you can measure. The graveyard in that segment is now longer than the pilot list.

5. The first real embodied-AI revenue went to carmakers

• Summary: Unitree rose 16% in a session and its market cap returned to ¥200 billion. In the same stretch, Shenlan Robotics was in bankruptcy proceedings with wages unpaid. TMTPost's reading: the companies booking actual embodied-AI revenue are largely automakers and their suppliers, while humanoid makers book valuations.
• Source: TMTPost · 2026-10-28
• Editor's Take: One company's market cap and another's bankruptcy filing, in the same week, in the same sector. That is what a hardware shakeout looks like before volume arrives.

---

II. Autonomous Driving

1. Volkswagen puts L3 at 2027 and is designing its own L4-grade chip

• Summary: At its first China technology day, Volkswagen Group China said it plans to have L3 functions ready by 2027 and confirmed it is developing C7H, a system-on-chip built to support up to L4. The chip work sits alongside the group's VEGA platform and electronics architecture.
• Source: ITHome · 2026-10-28
• Editor's Take: A 2027 L3 date from a legacy OEM is a procurement schedule, not a product promise. The chip matters more. Whoever owns the SoC owns the update cadence for the next decade of that car line.

2. Three Zunjie V800s snapped their brake pedal brackets in the same test

• Summary: Dongchedi's test video, published October 8, showed three new Zunjie V800s in 100 km/h emergency stops. The pedal bracket failed on the second, third and fourth full presses, at the same point, on cars from different production batches. The case is being read as a test of Huawei's car BU, which is trying to become the Bosch of Chinese auto.
• Source: Huxiu · 2026-10-28
• Editor's Take: A pedal bracket is a cheap stamped part with a ten-year liability tail. Three batches failing identically puts the supplier qualification process under review, not the driver's foot.

3. A Yijing X9 caught fire after a rear-end hit. Three occupants walked away.

• Summary: Video of a Yijing X9 burning in Taizhou, Zhejiang circulated on October 28. Brand general manager Zeng Qinglin said the car had already braked to a stop when a loaded dump truck hit it from behind, and that none of the three people inside were injured.
• Source: ITHome · 2026-10-28
• Editor's Take: EV fire footage gets read as a battery story. This one is a crash-structure story, and answering within hours is the part other OEMs should copy.

4. Volkswagen may ship its China-developed electrical architecture to Europe

• Summary: Reuters, citing Manager Magazin, reported that VW is weighing whether to put the E/E architecture currently used only in China into European mass-market models. The stack was developed with XPeng and CARIAD China. The argument for moving it west is cost.
• Source: Huxiu · 2026-10-28
• Editor's Take: The direction of technology transfer has flipped. For twenty years the platform came from Wolfsburg and was localized in China. Now the cheap, shipped software stack is Chinese, and Europe is the follower market.

---

III. World Models & Physical AI

1. Baidu is about to show a world model, aimed at robotaxis first

• Summary: Cailianshe reported that Baidu plans to disclose progress on a world model in the coming days, with autonomous driving as the first application. A driving world model predicts how a scene will evolve instead of only labelling what is in it, which is what a planner needs at an unprotected intersection.
• Source: ITHome · 2026-10-28
• Editor's Take: Baidu's edge here is operational data from Apollo's live zones, which is precisely what a driving world model needs and precisely what cannot be bought.

2. FluxVLA Engine wants embodied teams to stop rebuilding their plumbing

• Summary: Leiphone walks through the technical report behind FluxVLA Engine, an open engineering stack for vision-language-action models. The pitch is that teams should reuse the data pipeline, training harness and deployment tooling, and spend their engineers on the model instead.
• Source: Leiphone · 2026-10-28
• Editor's Take: Every embodied startup rewrites the same data loader and the same evaluation rig, then wonders why it ships slower than a lab with a platform group. Shared infrastructure is the least glamorous fix and the one that pays back fastest.

3. A world-model startup founded in December took four rounds from Sequoia China in months

• Summary: LiberAI, founded in December 2025 and working on physical-world models, took a Sequoia China angel round in March 2026 and four follow-on rounds within months. The company came out of a Tsinghua cohort. The article's frame is that the entry ticket for this category has changed.
• Source: Huxiu · 2026-10-28
• Editor's Take: Stacking four rounds into a seed-stage company is a supply-constrained bet. There are perhaps a few dozen people in China who have trained a world model at scale, and the money is chasing them, not the idea.

4. WIKA is selling embodied AI to construction sites, one dataset at a time

• Summary: WIKA Mobile Control spent the second half of 2026 at Huawei Connect and Alibaba's Yunqi conference pitching datasets that connect construction machinery to embodied models. The target is machine autonomy in an industry where the operator is the bottleneck and the site is already fenced.
• Source: QbitAI · 2026-10-28
• Editor's Take: Construction is a better first market than the home. The site is controlled, the tasks repeat, and a machine that runs a full shift without a driver pays for itself in one season.

---

IV. Foundation Models

1. Artificial Analysis rebuilt its index, and 60% of the weight now sits on hidden answers

• Summary: Intelligence Index v5 is the largest rebuild since v4 landed in January 2026. 60% of the evaluation weight now rests on private or answer-hidden test sets, a direct response to benchmark contamination. On the new FrontierCode 2.0, DeepSeek's V4.1 Flash 1017 is the highest-scoring open-weight model.
• Source: ITHome · 2026-10-29
• Editor's Take: Contamination was the open secret behind half the launches this year. Hiding the answers does not end the arms race, it moves the advantage to whoever can afford fresh evaluation data.

2. DeepSeek V5 identifiers surface in the Harness repository

• Summary: Developers found three previously unpublished model strings, deepseek-v5, deepseek-v5-open and deepseek-v5-next, in the provider spec test file of the official DeepSeek Harness repo, around line 154.
• Source: ITHome · 2026-10-28
• Editor's Take: A test fixture is a weak signal and cheap to plant. The useful read is that DeepSeek now ships a harness, meaning an agent runtime alongside the weights, which is where the lock-in actually lives.

3. Kimi K3.1 Pro cut API prices by a third, then hit a token-bill problem

• Summary: Moonshot released K3.1 Pro on October 24 with flagship-tier claims and API prices roughly one third below the previous generation. Users then found that token consumption per task ate the discount, and the price-per-useful-answer argument fell apart on social feeds.
• Source: Huxiu · 2026-10-28
• Editor's Take: Cutting the price per token while raising tokens per task is a margin transfer, not a price cut. Buyers have started pricing the outcome, and that is bad news for anyone selling on sticker price.

4. Ant open-sources Ling-3.1-flash-1019 for personal agents

• Summary: Ant Group's Ling team released Ling-3.1-flash-1019 as open weights on October 28, positioned for personal-agent workloads rather than chat. The pitch is tool use and multi-step tasks at a cost that lets an assistant run all day without the bill becoming the product decision.
• Source: QbitAI · 2026-10-28
• Editor's Take: Personal agents live or die on cost per hour of attention. A cheap open flash model aimed at that job is a bet that the assistant layer, not the frontier, is where Chinese labs can win.

5. Moore Threads open-sources a training stack for 10,000-card MoE runs

• Summary: Moore Threads released Megatron-MUSA-Titan, a training infrastructure project built on its own full-function GPUs, targeting trillion-parameter MoE training on clusters of 10,000 cards and covering both pretraining and reinforcement learning.
• Source: ITHome · 2026-10-28
• Editor's Take: Software is where Chinese GPU vendors kept losing. Open-sourcing the trainer gets other people to debug your interconnect for free, which is the fastest route to being taken seriously by model teams.

6. A Chinese inference startup took 88 first places on Nvidia's own kernel benchmark

• Summary: On SOL-ExecBench, an NVIDIA Research benchmark for GPU operator performance, Infinigence placed first in 88 categories. Part of the work involved agents participating in low-level kernel optimization rather than humans tuning by hand.
• Source: QbitAI · 2026-10-28
• Editor's Take: Kernel work was supposed to be the moat around CUDA. A team using agents to search the optimization space suggests that at least part of that moat is software-addressable now.

7. Tencent open-sources a simultaneous-translation model with a 3-second lag

• Summary: Tencent released Hunyuan Hy-SimulMT, an open-source simultaneous machine translation model that the company says holds accuracy while keeping average response time to about 3 seconds.
• Source: TMTPost · 2026-10-28
• Editor's Take: Three seconds is roughly the width of a spoken clause. Below that, simultaneous translation feels like subtitles. Above it, listeners lose the thread of the meeting.

---

V. AI Software

1. Doubao Work will record your desktop routine and replay it as a skill

• Summary: ByteDance's Doubao Work upgraded its PC-control feature with better click precision and speed, and added the ability to record a demonstration of a desktop task and save it as a reusable skill. Users can narrate while they demonstrate.
• Source: ITHome · 2026-10-28
• Editor's Take: Recording a workflow is the cheapest form of programming ever handed to an office worker. The failure mode is the one macros had in 1998: the skill breaks the moment an app moves a button.

2. Tencent's WorkBuddy lets a phone take over the desktop

• Summary: WorkBuddy added remote control, built with support from Tencent's Edge One team. The phone shows a live view of the desktop and accepts direct input, so a user can watch a running task instead of reading its output after the fact.
• Source: ITHome · 2026-10-28
• Editor's Take: Agent products keep rediscovering one thing: people do not trust what they cannot watch. A live screen is a trust feature wearing the costume of a convenience feature.

3. Two Shanghai personal-agent startups raised within weeks of each other

• Summary: Today AI, part of Huihe Intelligence in Shanghai's Xuhui district, closed an angel round above $200 million for a software-only personal agent. Flowtica AI raised tens of millions of yuan for a magnetic clip-on device in the same category.
• Source: Huxiu · 2026-10-28
• Editor's Take: A $200 million angel round for an assistant with no hardware is a bet on distribution, not technology. The real contest is which app gets permission to act on your behalf, and that is a platform fight.

4. Beijing published an 18-point plan on AI and employment

• Summary: The State Council General Office released a document on adapting to AI's effect on jobs, with 18 tasks including softening displacement, building employment-risk assessment, and adjusting income distribution and social security.
• Source: Huxiu · 2026-10-28
• Editor's Take: The risk-assessment mechanism is the item to watch. It makes AI deployment inside Chinese firms reviewable before it happens, which no other major economy has put in writing.

5. Lao Gan Ma's AI quality-control boast became a brand crisis

• Summary: A September 29 provincial newspaper piece quoted a Lao Gan Ma executive saying its AI visual inspection system was fully live, raw materials 100% traceable, and the defect rate down to 0.02%. Readers turned the numbers into a question about what the company had been shipping before.
• Source: Huxiu · 2026-10-28
• Editor's Take: Consumer brands keep learning that an efficiency metric doubles as a confession. Announcing a 0.02% defect rate invites buyers to ask what it used to be.

6. The county-town AI project that ate ¥700,000 and delivered a dashboard

• Summary: A software firm with fewer than a hundred staff in a fifth-tier Shaanxi city spent about ¥700,000 from August on a system pitched as lead generation, customer service, recruiting and decision support. What arrived, per the account, was a set of demos that never touched the actual workflow.
• Source: Huxiu · 2026-10-28
• Editor's Take: This is the forward-deployed-engineer model without the engineer. Vendors sold an outcome, shipped a dashboard, and left the integration to a company with nobody to do it.

7. Kingdee wants ERP to become the memory layer for agents

• Summary: Kingdee said its Lingji agent platform has 3,706 paying customers, with 26 signed at launch and more than 40 agents live; AI-native product revenue grew 189% year on year. The positioning is ERP as the context base an agent reads before it acts.
• Source: TMTPost · 2026-10-28
• Editor's Take: ERP vendors hold the one asset agent startups cannot buy, which is twenty years of how a specific company actually runs. Whether they can expose it without breaking their own permission model is the open question.

---

物界前沿 | EastVoice