--- date: 2026-09-21 subject: "RF-spying satellites launch | US-China AI talks | Zhipu code leak" --- **China puts three-satellite radio-frequency collection formation into orbit** yesterday, one of nine satellites aboard, giving its operator a way to map the electromagnetic spectrum and locate radio transmitters from space. **Vice Premier He Lifeng met with U.S. representatives** Treasury Secretary Scott Bessent and Trade Representative Jamieson Greer in New York yesterday for trade talks that also covered AI, ahead of Xi Jinping's summit with Trump in Washington this week. **Zhipu open-sourced its ZCode coding assistant** today and apologized after an independent developer found the client uploading encrypted snapshots of signed-in users' workspaces to Alibaba Cloud storage, and a Taiyuan software firm said more than 30,000 of its files were uploaded. # Top Stories - **Chinese launch orbits the country's first commercial three-satellite radio-frequency collection formation** — A Lijian-1 rocket built by commercial firm CAS Space placed nine satellites in orbit yesterday from the Dongfeng commercial space zone in northwestern China, among them the Diting-01 A, B and C satellites. Those three fly as a formation collecting radio-frequency data from orbit, which Science and Technology Daily calls a first for a Chinese commercial constellation. The formation gives wide-area awareness of the electromagnetic spectrum and can fix the position of a radio transmitter quickly and precisely. Tianjin Xingkan Jiuzhou Technology built the satellites. Also aboard was Pengcheng Nanke-1, from GalaxySpace, Peng Cheng Laboratory and Southern University of Science and Technology, which will test satellite-to-ground laser links at 100 gigabits a second. This was the 16th flight of the Lijian-1. [Science and Technology Daily](https://epaper.stdaily.com/statics/technology-site/index.html#/home?isDetail=1¤tNewsId=4b82233c46e242729cbd8a281438ba90¤tVersionName=02%C2%B7%E7%AC%AC02%E7%89%88%EF%BC%9A%E8%A6%81%20%E9%97%BB¤tVersion=2&timeValue=2026-09-21) [state media] [China Military Online](http://www.81.cn/ss_208539/16487786.html) [state media] - **China's lead trade negotiator discussed ai in talks with Bessent and Greer** — Vice Premier He Lifeng, a Politburo member and China's lead negotiator with Washington, met Treasury Secretary Scott Bessent and U.S. Trade Representative Jamieson Greer in New York yesterday. The Xinhua readout lists a dialogue on AI issues alongside the economic and trade consultations rather than inside them, a separate track in the official Chinese account. On the trade side, it describes candid exchanges on implementing the consensus reached in previous rounds and on promoting bilateral trade and investment. It frames the talks as guided by the two presidents' consensus and the principles of mutual respect, peaceful coexistence and win-win cooperation. [Caixin](https://international.caixin.com/2026-09-21/102486990.html) - **Zhipu open-sources its coding assistant after a client says it uploaded more than 30,000 files** — Zhipu said today that it has finished fixing its ZCode coding assistant, apologized to all users and published the client's source code for community inspection. An independent developer found Friday that ZCode builds an encrypted snapshot of a signed-in user's workspace, complete version-control history included, and sends it to Alibaba Cloud storage. Zhipu attributed that to a repository-indexing feature that shipped switched on by default, supporting session restore, version rollback and automatic documentation. Chengming Technology, a software firm in Taiyuan, says its forensic analysis found six workspaces uploaded, the largest holding 32,932 files, and that the material included database passwords, interface keys and unreleased business plans. Indexing a code base can be done on the user's own machine by hashing files, so uploading a full workspace is generally unnecessary, iiMedia Research chief executive Zhang Yi said. Zhipu will invite the China Academy of Information and Communications Technology, the state telecom research institute, and security vendor NSFOCUS to audit the client. [Huxiu](https://www.huxiu.com/article/4892857.html) - **StepFun prices its new flagship model at one-eighth the per-task cost of Claude Opus 5** — StepFun released Step 5 Preview yesterday, a flagship model aimed at agent work where a system carries out a task over many steps rather than answering in a single reply. StepFun puts the model's per-task cost at one-eighth that of Anthropic's Claude Opus 5. It holds 600 billion parameters but runs only 27 billion of them for any given token, which is what keeps the serving cost down. Artificial Analysis, an independent benchmarking service, rates the model well priced against its peers but describes it as verbose. It generates far more tokens per task than the median model, which eats into the low headline price on long answers, AI/TLDR notes. Step 5 Preview is proprietary and served only through StepFun's programming interface, while PingWest reports the company will publish full weights on Oct. 15. [PingWest](https://www.pingwest.com/w/317594) [AI/TLDR](https://ai-tldr.dev/releases/stepfun-step-5-preview/) [Artificial Analysis](https://artificialanalysis.ai/models/step-5) # Policy & Regulation - **China's internet regulator drafts a ban on virtual companion services for minors** — A draft regulation from the Cyberspace Administration of China, which regulates online content and data, would bar providers from offering minors virtual intimate relationship services such as virtual kin and virtual partners, along with social networking with strangers. Providers would also be barred from deploying algorithm models that induce emotional dependence, addiction or excessive spending in minor users, and would have to run review systems covering those algorithms. The draft separately bars livestream publishing services for anyone under 16, and requires providers to be able to identify which of their users are minors. Public comments close Oct. 17. [China Military Online](http://www.81.cn/ss_208539/16487411.html) [state media] # AI & Foundation Models - **CXMT mass-produces fifth-generation DRAM using quadruple patterning** — ChangXin Memory Technologies said its fifth-generation memory platform has entered mass production, disclosing the step yesterday at the World Manufacturing Convention in its home city of Hefei. The platform reaches an 11.95-nanometer active-area half-pitch in its memory arrays using quadruple patterning, a process that repeats several manufacturing steps to produce finer circuit patterns. DRAM is the working memory that holds whatever a processor is using at that moment. CXMT says the platform produces at least 50% more chip dies per wafer than its fourth-generation one, a measure of how densely dies pack onto the wafer rather than a count of working chips. Two 24-gigabit LPDDR5X parts built on the platform are already in volume production for smartphones. Vice President Luo Xiaodong told the convention that "our process capability is now on par with the most advanced mass-produced nodes out there in the industry." [Leiphone](https://www.leiphone.com/category/zaobao/W5fLfeynETtTWpTt.html) [TechNode](https://technode.com/2026/09/21/cxmt-announces-mass-production-of-fifth-generation-dram-platform/) [Yahoo Finance](https://ca.finance.yahoo.com/news/chinas-cxmt-says-memory-chip-035403310.html) - **Qinghai lab releases what it calls China's first Tibetan multilingual AI input method** — The State Key Laboratory of Tibetan Intelligence at Qinghai Normal University released the Zhida input method yesterday, along with a matching Tibetan font set. An input method is the software that turns keystrokes, speech or photographs into written characters. Rather than just translating, it lets people write Tibetan on phones and computers by typing, speaking or photographing text, and uses the lab's AI model to correct spelling and grammar, turn spoken phrasing into written language and translate in both directions as they write.Behind it sits the laboratory's Zhida large model, released in nine versions running from 800 million to 100 billion parameters and supporting translation among 56 languages. [Sina Tech](https://finance.sina.com.cn/tech/digi/2026-09-21/doc-inispuef2169329.shtml) [International Science and Technology Innovation Center](https://www.ncsti.gov.cn/kjdt/kjrd/202609/t20260921_256853.html) [govt] - **Third-party developers port NetEase Youdao's speech models into local inference tools** — A third-party port of NetEase Youdao's Confucius4-R2T2 speech-recognition model was merged Saturday into audio.cpp, an open-source framework for running audio models on a personal computer. That moves the model out of Youdao's own serving stack and into tools developers run on their own machines. R2T2 transcribes speech as it arrives rather than waiting for a complete utterance, and its companion Confucius4-T3PO, a 14-billion-parameter model, translates the same way. Both topped their category trending lists on Hugging Face, R2T2 for speech recognition and T3PO for translation. Hugging Face is the main repository developers use to publish and download open model weights. Overseas quantization developers have since published compressed builds of both models, and the R2T2 builds can be called through the llama.cpp and Ollama local-inference runtimes. [PingWest](https://www.pingwest.com/w/317657) # Robotics & Autonomous Systems - **Lexiang Technology claims a robot model that keeps learning after deployment** — Lexiang Technology says its Aether model keeps updating while it runs on a robot, and calls it the first self-evolving embodied intelligence model. The company staked that claim on an unscripted outdoor livestream. Robots running Aether took random food orders and worked a barbecue, handling ingredients and tools while coping with interruptions. The robots did some of the barbecue work autonomously, with pauses, failures and readjustments along the way. Two robot bodies traded task information and divided the work. GeekPark writes that Lexiang has so far offered only its own account of the mechanism, leaving open how much the model weights actually change. Aether has accumulated experience across about five robot bodies spanning dual-arm, biped and wheeled forms. Its control loop reads state and issues joint commands more than 200 times a second. [GeekPark](https://geekpark.net/news/370665) - **Tsinghua-led team open-sources a layer that lets general AI models drive robot arms** — Tsinghua University, Infinigence AI and Zhengxing Innovation published RPent today, an open-source framework that sits between a general large model and a robot. It hands planning and task interpretation to the general model while leaving fine manipulation to vision-language-action models, which translate camera images and language instructions directly into joint commands. On LIBERO-PRO, a standard suite of simulated manipulation tasks, the team reports a 92.6% success rate with GPT-6 Astra doing the planning. Adding RPent's memory mechanism raised the success rate on one disturbance task, in which objects swap position partway through, from 31% to 87%. In demonstrations, RPent-driven robots poured steel balls into a bowl and moved two bowls aside to reach a spoon hidden underneath. [QbitAI](https://www.qbitai.com/2026/09/493218.html) - **Genisom AI releases software letting several legged robots carry one load together** — Genisom AI's PITA brand released a multi-legged collaborative transport system Saturday, in which several legged robots cooperatively carry a single payload through an environment they have not mapped in advance. Holding a shared load is harder than walking in formation, because the machines are physically coupled through the object they are carrying. An operator wearing a virtual-reality headset designates a target location from a safe distance. The Gangbeng L1-W EDU robots then perceive the surroundings, locate one another and share spatial information to plan the route. Facing pedestrians, low obstacles and stairs, they adjust formation and path in real time according to passage space and load state. Genisom AI names casualty evacuation in earthquake rubble, handling of hazardous chemical and radiation materials, and moving heavy parts across industrial parks as target applications. [Gasgoo Intelligent Vehicles](https://auto.gasgoo.com/news/202609/20I70472636C601.shtml) - **PrimeBOT puts two personal robots on sale from 19,999 yuan** — PrimeBOT began selling its Q1 and T1 personal robots yesterday at a Shanghai launch, both starting at 19,999 yuan, about $3,000. That pricepoint puts a walking robot in consumer-electronics price territory, with deliveries beginning Oct. 1 in order sequence. A T1 Pro version adds an Nvidia Orin Nano compute module, 360-degree obstacle avoidance and automatic return-to-charge navigation, and opens interfaces to outside developers. PrimeBOT plans to spend 500 million yuan, about $75 million, on research and development during 2026. The company says its line turns out one robot every 2.5 minutes. [PingWest](https://www.pingwest.com/w/317608) # Space - **China says its Wangyu lunar spacesuit has cleared dozens of key technologies** — Development of the Wangyu lunar spacesuit has broken through dozens of key technologies, officials said Saturday at the first Aerospace Medical Engineering Conference in Tianjin. Extravehicular suits are the pressurized garments that let an astronaut work outside a spacecraft. Only a handful of countries can develop them independently, the report says. Zhang Wanxin of the China Astronaut Research and Training Center said the in-orbit working life of Chinese suits now stands at four years of storage with no fewer than 20 spacewalk uses in that span. Chinese suits have also supported a single spacewalk of nine hours, which the center calls a world record. Work on the Wangyu suit began in 2020, and China is targeting a first crewed lunar landing before 2030. [China Military Online](http://www.81.cn/kt/16487599.html) [state media] - **China finishes a bed rest study behind year-long spaceflight and lunar missions** — The China Astronaut Research and Training Center has completed its Dixing-3 bed rest experiment, the Tianjin conference heard Saturday. The center says the results show good protection of the bone-muscle and cardiovascular systems. Bed rest substitutes for spaceflight on the ground. Qu Lina of the center said a head-down tilt shifts body fluid toward the head and chest and unloads the leg bones and muscles, closely resembling what astronauts undergo in weightlessness. Volunteers held either a 6-degree head-down or a head-up posture for 15 or 30 days, eating, washing and sleeping in that position, and some performed prescribed daily exercise. The design validated two countermeasure packages at once, one for a year-long space-station flight and one for the reduced gravity of a crewed lunar mission. One member of the Shenzhou-23 crew will carry out a year-long in-orbit residency. [China Military Online](http://www.81.cn/kt/16487613.html) [state media] # Strategic Context - **China's foreign ministry refers rare-earth resumption questions to other agencies** — Foreign ministry spokesperson Guo Jiakun declined to say whether China has carried out the rare-earth supply resumption agreed at an earlier meeting, telling reporters that specific questions should go to the competent Chinese authorities. He gave the same answer when asked whether Beijing would extend the Nov. 10 deadline on its expanded rare-earth export restrictions, and by how long. Guo said China is consistently committed to maintaining the stability and security of global critical-minerals supply chains, and that both sides should implement the consensus reached by the two presidents. The questions came from a New York Times reporter, following a U.S. official's statement last week that China has not fully implemented the resumption. [Wallstreetcn](https://wallstreetcn.com/livenews/3168067)