--- date: 2026-08-24 subject: "Robot breaks human track record | Xiaomi's 3nm driving chip | CAC sets five-year tech targets " --- **A bi-pedal humanoid robot ran the 100 meter dash in 9.32 seconds** at the second World Humanoid Robot Games, which opened Aug. 22 in Beijing, beating Usain Bolt's 9.58-second human world record. **Xiaomi unveiled a 3-nanometer driving chip** today, the Xring D100, aimed at reducing the company's reliance Nvidia Thor chipset, with commercial use scheduled for 2027. **The Cyberspace Administration published an action plan** on Friday covering 2026 through 2030 that steers network and cloud firms toward high-end chips and training clusters, and calls for faster onshore listing review. # Top Stories - **Humanoid robots break the human 100-meter record at Beijing games** — The second World Humanoid Robot Games opened Aug. 22 at the National Speed Skating Oval in Beijing, People's Daily said. Entries total 2,056 robots from 666 teams in 16 countries across 51 events, making the competition a public benchmark for Chinese humanoid locomotion hardware. The Communist Party newspaper measured the results against human track records rather than against other robots. In the 100-meter dash preliminaries for the large-robot class, a machine from the Tianzhuo team ran 9.39 seconds, beating Usain Bolt's 9.58 world record. Later heats produced 9.34 seconds from a joint Tiangong-Jingfeng Technology entry and 9.32 seconds from both the Tianxiao and Tiangong teams. In a standing high jump demonstration, the Beijing-built Tiangong Ultra cleared 2.8843 meters, roughly 9.5 feet. [People's Daily](https://paper.people.com.cn/rmrb/pad/content/202608/23/content_30176540.html) [state media] - **Xiaomi launches a 3-nanometer in-house driving chip aimed at reducing Nvidia reliance** — Xiaomi unveiled three in-house Xring chips today, among them the D100, a smart-driving processor the company says has completed validation and is scheduled for commercial use in 2027. Its electric vehicles rely mainly on Nvidia Thor parts for smart driving, so the D100 targets a component a U.S. supplier now holds. Xiaomi describes the D100 and the Xring O3 phone processor as designed on a 3-nanometer process, among the most advanced nodes in commercial production. The D100 supports up to 160 gigabytes of unified memory, which Xiaomi says is enough to run a 200-billion-parameter language model on the vehicle rather than in a data center. CCTV News, the state broadcaster, posted on its official Weibo account that the simultaneous launch was "another breakthrough for China's chip industry." [IT Home](https://www.ithome.com/0/993/602.htm) [CnEVPost](https://cnevpost.com/2026/08/24/xiaomi-unveils-xring-d100-smart-driving-chip/) - **China's top cyberspace body names AI chips and training clusters as priorities through 2030** — The Cyberspace Administration of China published an action plan on Friday covering 2026 through 2030 for the firms that build and run the country's networks, cloud platforms and data infrastructure. The plan pairs technology targets with changes to the financing those firms can reach, and it names the same areas U.S. export controls and investment screening cover. On technology, it directs enterprises toward high-end chips, foundational and industrial software, and the key technology behind high-performance training clusters, the linked machines a frontier model is trained on. Quantum technology, brain-computer interfaces and new-type displays are named as frontier targets. The capital markets provisions call for faster onshore listing review and encourage listings on the STAR Market and ChiNext, the Shanghai and Shenzhen boards for technology companies. The plan also calls for drafting a telecommunications law and a cybercrime prevention law. [CAC](https://www.cac.gov.cn/2026-08/21/c_1789061449039952.htm) [govt] - **Chinese filing rules put ByteDance and DeepSeek models in Teslas** — Volcano Engine, ByteDance's cloud unit, said on Aug. 19 that Tesla's in-vehicle system in China now runs the ByteDance large language model Doubao. It is the first time a third-party Chinese model has been built into Tesla's car system in its decade-plus in the market. The cars use a dual-model setup reached through Volcano Engine, with Doubao handling navigation, media and climate commands and DeepSeek Chat handling conversation and information queries. Tesla vehicles in the United States run xAI's Grok, which has not completed the filing the Cyberspace Administration of China requires before a generative AI service can operate in China. Separate rules require voice and location data collected by vehicles to be stored and processed inside the country. [Gasgoo](https://auto.gasgoo.com/news/202608/21I70470020C601.shtml) - **DeepSeek puts image input on its cheaper model, pricing below U.S. rivals** — DeepSeek released a vision model called V4-Flash-Vision-Exp on Friday and opened an image input API alongside it. The model runs on Flash, the cheaper of DeepSeek's two tiers, so developers get image handling without moving up to the larger Pro model. Flash carries 284 billion parameters in a mixture-of-experts design, an architecture that activates only a slice of its parameters on each query. An image consumes at most 384 tokens, the units models bill by, at standard Flash pricing. GPT and Claude typically use 800 to 1,100 tokens on images of comparable resolution, per GeekPark. In a scene-understanding test on release, with thinking mode on, the model spent its whole completion budget on reasoning and returned no output at all. Setting reasoning effort to none produced full output on the same test. [GeekPark](https://geekpark.net/news/369278) # Policy & Regulation - **China's industry ministry tells robot makers to weigh automation gains against jobs** — The Ministry of Industry and Information Technology (MIIT), which sets industrial policy and licenses telecom operators, laid out three priorities for robot policy at the World Robot Conference in Beijing on Aug. 19. Two cover joint research between Chinese and foreign institutions and firms, and application pairing in fields including agriculture, medical care, elder care and emergency response. The third is a people-oriented approach calling for development to be upward and good for the people while making sure efficiency gains are balanced against employment security. The ministry put first-half revenue for the robot industry at 165.5 billion yuan ($24 billion), up 24.5% year on year. Smaller robot startups with smaller revenue sit outside of the government figures. [MIIT](https://wap.miit.gov.cn/xwfb/bldhd/art/2026/art_18a2e30182d44a4ea7c2c1295704dfee.html) [govt] - **Draft Chinese standard would make cars verify their software cryptographically at every startup** — MIIT released a draft mandatory national standard for public comment on Friday, "Technical Requirements for Automotive Cryptography," prepared by the National Technical Committee of Auto Standardization. A mandatory national standard, CnBeta writes that it would move automotive software security from corporate self-discipline to compliance for every vehicle model sold in China. The core obligation is secure boot, the chain of signature checks a system runs before it starts. Firmware, bootloaders, operating systems and key application software that has been tampered with, replaced or forged must be blocked from running. The draft extends the same check to over-the-air update packages before they install. Testers verify compliance by replacing each of those software images in turn, powering the vehicle on and confirming that step-by-step verification runs at each stage. [CnBeta](https://www.cnbeta.com.tw/articles/tech/1574700.htm) # AI & Foundation Models - **Lightwheel AI open-sources dataset of human recordings for robot training** — Lightwheel AI open-sourced 100,000 hours of first-person human recordings on Thursday at the World Robot Conference in Beijing, a collection it calls EgoSuite-Open100K. Robots learn manipulation from video of human hands rather than from text, so opening a large tranche of that scarce material pulls outside developers onto Lightwheel AI tooling. The dataset spans more than 15,000 scene categories and many task types, with annotations covering hand pose, full-body pose and event-level semantics. The first batch is on Hugging Face for academic and commercial use. [PingWest](https://www.pingwest.com/w/316686) - **Testers trace an anonymous coding model on OpenRouter to China's Zhipu AI** — A model called Ox Alpha appeared on the OpenRouter marketplace on Thursday, listed only as coming from an anonymous provider. Within days, independent testers had assembled technical evidence pointing at Zhipu AI, the Beijing developer behind the GLM model family. In a widely circulated tokenizer test, one mixed Chinese-English passage encoded to 68 tokens under GLM-5.3 and GLM-5.2 but to 143 tokens under Ox Alpha. Testers read the extra tokens as an externally appended hidden system prompt rather than a different tokenizer. Not everyone accepts that reading: Google DeepMind researcher Evan Otero posted the word "Gemini" in the discussion, which the community took as a hint at a Google origin. OpenRouter confirmed after the fact that Pony Alpha, an anonymous model that surfaced there in February, was an early test version of GLM-5. [QbitAI](https://www.qbitai.com/2026/08/478191.html) - **Alibaba's DAMO Academy puts a liver metastasis detector into daily hospital use** — Alibaba's research arm, DAMO Academy, released a medical imaging model called DAMO LiON, built to catch primary liver cancer and the metastases radiologists most often miss. It extends DAMO's earlier cancer work from screening into diagnosis. Screening sorts a population; diagnosis guides one patient's treatment. Developed with Shengjing Hospital of China Medical University and published in Nature Medicine, the model read enhanced CT scans for more than 10,000 patients in a two-month trial. It helped doctors find 15 malignant tumors that had originally been missed, most of them lesions around a centimeter across, sending them to treatment sooner. DAMO algorithm expert Yan Ke said the lesions it flags are generally small, faint and in atypical positions. When the model's assessment conflicts with a doctor's reading, AIbase reports, the case goes to a senior physician for review. [PingWest](https://www.pingwest.com/w/316711) [AIbase](https://news.aibase.com/zh/news/30564) # Chips & Semiconductors - **GPU designer Moore Threads splits AI serving across two hardware pools to cut cost per token** — Moore Threads published a white paper on its MTT S5000 card, saying one machine can push 5.8 million tokens a minute in a long-context test. It is one of the domestic options Chinese AI firms are turning to under U.S. export controls on Nvidia hardware. Serving a model runs in two stages: prefill reads the prompt and is limited by raw compute, while decode writes the answer token by token and is limited by memory bandwidth. Running both on one pool wastes capacity in both directions, per Sina Finance, so each stage gets the hardware that suits it. IT Home reports one pool tuned for the shortest time to first token, the other for many simultaneous users. A fully loaded machine can earn up to 646,000 yuan ($95,000) a month for whoever rents it out. [PingWest](https://www.pingwest.com/w/316710) [IT Home](https://m.ithome.com/html/993370.htm) [Sina Finance](https://finance.sina.com.cn/tech/digi/2026-08-24/doc-inipkrka2998647.shtml) - **Zhongcheng Hualong releases inference chip and a 10,000-card cluster design** — Zhongcheng Hualong introduced its second-generation inference chip, the HL200, in Beijing, alongside a cluster design it says scales to more than 10,000 cards. Inference is what China most needs to substitute at home, since running models in production consumes far more silicon over time than training does. The HL200 natively supports FP4 and FP8, low-precision formats that trade numerical accuracy for speed. The company says the chip reaches 4 petaflops per card at FP4, or 4 thousand trillion calculations a second. MIIT committee official Bi Kaichun praised the company for a fully domestic tape-out, a design carried to manufacturing without foreign steps. Zhongcheng Hualong's products have won bids with China Mobile and China Telecom covering more than 20 provinces. [QbitAI](https://www.qbitai.com/2026/08/478469.html) # Robotics & Autonomous Systems - **Pony.ai and Uber plan more than 2,000 robotaxis in five European cities** — Pony.ai and Uber said on Aug. 14 that they will deploy more than 2,000 robotaxis across five European cities. CTD's [August 11th edition](https://chinatechdaily.org/archive/china-daily/2026-08-11/) covered Pony.ai's 100-million-kilometer autonomous driving milestone. The arrangement puts Chinese driving software inside an American demand platform, with vehicles owned by different partners depending on the market, TechCrunch reports. Pony.ai already runs robotaxi service in several Chinese cities on L4 software, the level at which a vehicle drives itself inside a defined area with no human backup. The expansion builds on the commercial service the two companies run in Zagreb, Croatia, which will soon be plugged into the Uber app. Its seventh-generation robotaxi has reached vehicle-grade mass production capability and has turned unit economics positive in some Chinese cities, Pony.ai says. [EV Industry News](https://www.evchanye.com/2026/0814/51080.shtml) [TechCrunch](https://techcrunch.com/2026/08/14/uber-and-pony-ai-plan-to-bring-2000-robotaxis-to-europe/) - **Chinese robot makers report six-to-eight-hour runtimes, up from about an hour** — Multiple exhibitors at the World Robot Conference, which opened in Beijing on Aug. 19, said their robots now run six to eight hours on a full charge, against roughly one hour previously. That covers a standard work shift, long enough for a machine to hold a job rather than run a demonstration. Spirit AI's Moz1 was told to tidy a living room and broke the job into subtasks by itself, adjusting its route when staff moved objects mid-demonstration. Industry figures at the show said pairing world models with vision-language-action systems, which turn camera input into robot motion, has decoupled a model from any one robot body. A world model learns how an environment behaves in order to predict what happens next. Moving a model to a new deployment scene now takes about a week rather than two months, they said. [Huxiu](https://www.huxiu.com/article/4885514.html) ```markdown - **Beijing Humanoid debuts humanoid robot and a Mercedes-Benz dealership deal** — Beijing Humanoid used the World Robot Conference in Beijing, which opened Aug. 19, to debut Tiangong Omni. The humanoid stands 1.35 meters tall and weighs 39 kilograms, or about 86 pounds. The company also announced what it calls a deep strategic partnership with Mercedes-Benz, putting a Chinese humanoid into a German automaker's customer-facing operations rather than a research lab. A different machine, Tiangong 3.0, worked the booth as a greeter, planning its own route around the crowd. Beijing Humanoid plans to put Tiangong 3.0 into Mercedes-Benz dealerships for guided tours, vehicle presentations and delivery ceremonies. [Gasgoo](https://auto.gasgoo.com/news/202608/20I70469784C601.shtml) - **Momenta clears the top automotive functional safety grade with a Chinese-Western driving stack** — A three-company autonomous driving platform announced on Aug. 19 has passed TUV Rheinland certification at ASIL D, the highest functional safety level the ISO 26262 automotive standard defines. Certification usually covers the pieces separately, and regulators increasingly want evidence at the level of the whole system. For Chinese suppliers selling driver-assistance technology abroad, that is the gate. Momenta, the Chinese self-driving software developer listed in Hong Kong, contributes the driving algorithms. XHEART supplies the vehicle-grade X7 processor that runs them, and BlackBerry's QNX unit supplies a safety operating system built on its SDP 8.0 release. The platform also meets Europe's UN R171 rule for driver-control-assistance systems, giving automakers a compliance basis for selling intelligent driving models internationally. The three companies describe the platform as open to mainstream global automakers rather than tied to one customer. [Gasgoo](https://auto.gasgoo.com/news/202608/19I70469660C601.shtml) - **Driverless airport tractors cut towing costs below 1 yuan per kilometer at Urumqi** — Driverless battery-powered tractors from UISEE Technology have cut the cost of towing cargo around Urumqi's Tianshan Airport below 1 yuan (15 cents) per kilometer, from about 2.2 yuan (32 cents) with fuel-powered tractors. Cost per kilometer is the number that decides whether driverless logistics spreads beyond demonstrations, in China or anywhere else. Airports suit the technology because the roads are private, the routes repeat and the labor is expensive. Nearly 100 unmanned vehicles now work at Tianshan, where driverless tractors handle more than 90% of cargo and mail transfer with no at-fault accidents. Co-founder and chief executive Wu Gansha estimated that driverless penetration among airport vehicles overall remains in the low single digits. Xinjiang Airport Group and UISEE plan to extend the service to all of Xinjiang's operating airports. [Gasgoo](https://auto.gasgoo.com/news/202608/14I70469253C601.shtml) - **Lidar maker Hesai shows a chip that senses color and distance together** — Hesai Technology showcased a chip platform called Picasso to the World Robot Conference in Beijing, calling it the first lidar chip to combine color sensing with distance measurement. Lidar maps a scene by timing laser pulses and normally needs a separate camera for color, so putting both on one chip aligns color and depth pixel by pixel. Hesai built its business on automotive lidar and used the show, which ran Aug. 19 to 23, to pitch a lineup for robots rather than cars. The ETX, the first lidar built on Picasso, reaches a maximum range of 600 meters, roughly a third of a mile. It supports up to 4,320 scan lines, a measure of how finely it slices a scene vertically. More than 50 companies including Unitree and Galbot are adopting its mini lidar, the JT128, Hesai says. [Gasgoo](https://auto.gasgoo.com/news/202608/20I70469789C601.shtml) # Strategic Context - **Chinese foreign ministry answers U.S. warnings on Chinese tech suppliers in Latin America** — Foreign ministry spokesperson Lin Jian said at a regular press briefing today that recent social media remarks by individual U.S. officials attacking Chinese technology companies fabricate facts and confound right and wrong. A reporter's question laid out the U.S. position, that high-risk technology from the Chinese companies concerned would create risks of espionage, cyberattacks and network disruption. Those officials also said U.S. partner countries in the Western Hemisphere should shift quickly to trustworthy suppliers to protect their citizens and sovereignty. Lin called the remarks tech bullying and unilateral coercion, and said Chinese technology companies are trustworthy suppliers that comply with the laws of the countries they operate in. The United States, he said, has long conducted surveillance, monitoring and cyberattacks in Latin America and the Caribbean. Lin urged Washington to stop interfering with countries' right to choose their own partners. [Sina Finance](https://finance.sina.com.cn/jjxw/2026-08-24/doc-inipkzxs2031370.shtml) [Wallstreetcn](https://wallstreetcn.com/livenews/3153743) [state-affiliated]