|
|
|
|
Bottom Line Up Front LandSpace launched 10 satellites today on a single Zhuque-2E rocket, described as the first large-scale constellation networking mission flown by a Chinese private launch company. A state newspaper commentary rejected a Reuters investigation that concluded Chinese military institutions are testing humanoid robots for battlefield use, arguing the report spliced together procurement notices, university papers and patents into a war-preparation picture none of them supports alone. The foreign ministry dismissed U.S. warnings that Chinese AI progress creates a national security problem, saying threat narratives and malign competition disrupt global AI governance and serve no party's interests. |
Rocket maker LandSpace flies the first private Chinese mission to build a satellite internet networkLandSpace put 10 satellites into orbit on a single rocket today, the launch foreshadowed by the rollout photos in the September 8th edition. The report calls it the first large-scale constellation networking mission flown by a Chinese private launch company. The Zhuque-2E vehicle, fueled by liquid oxygen and methane, lifted off from the Dongfeng Commercial Space Innovation Test Zone and released its payload at roughly 900 kilometers. Qianfan is the low-orbit broadband constellation run by Shanghai Yuanxin Satellite Technology, also called G60 Starlink, and this was its latest batch into polar orbit. Yuanxin system-department general manager Shang Lin said Sunday that the operator plans 324 satellites in orbit by the end of this year, the point at which it says regional service begins. |
| Chinese state newspaper disputes a Reuters finding that China's military is testing combat humanoids Science and Technology Daily published a commentary rejecting a Reuters investigation that concluded Chinese military institutions are testing humanoid robots for battlefield use. The dispute is over the evidentiary standard for treating China's humanoid robot industry as a military program The column argues the Reuters piece acknowledges on its own terms that no evidence shows any People's Liberation Army combat unit has deployed armed humanoid robots. It places the reporting in the threat-inflation pattern, arguing that Reuters spliced open procurement notices, university papers and corporate patents into a war-preparation picture none of them supports alone. Zhang's answer to who is actually pushing humanoids toward combat is the West, pointing to American reporting on a U.S. startup's robots sent to Ukraine and on military contracts covering weapons handling, while citing Western analysts who say the machines are years from being practical battlefield systems. |
| China's foreign ministry rejects U.S. warnings that its AI progress is a security risk Foreign ministry spokesperson Guo Jiakun was asked at a regular press briefing about the argument that Chinese progress in AI creates a national security problem for the United States. The question followed calls by Anthropic's Dario Amodei, OpenAI's Sam Altman and Elon Musk for AI companies to slow work on frontier systems. Amodei had said a continued Chinese lead in AI would pose a risk to U.S. national security. Guo said AI development concerns the common welfare of all humanity, and that all parties should jointly promote AI that is open, inclusive, broadly beneficial and oriented toward good. Spreading threat narratives and pursuing confrontation and malign competition, he said, disrupts the global AI governance process and serves no party's interests. |
| Shanghai launches a program to build a gigawatt-scale computing network in orbit Shanghai formally launched its Mingzhu Constellation Program on Sunday at a space-compute forum, targeting 2030 for what officials call an independently controllable space intelligent compute network. Satellites would carry the processors themselves rather than relay data to computers on the ground, processing what is generated in space before anything is sent down. The long-term goal is a gigawatt-class orbital computing constellation, alongside integrated space-ground compute networks and space energy technology. Independent controllability is a standing Chinese policy requirement that a system be buildable and maintainable on domestic technology, which shapes which chips the program can plan around. At the forum, the lead company signed industrial cooperation agreements with 27 supply chain partners, among them the Chinese Academy of Sciences' Innovation Academy for Microsatellites and Alibaba Cloud. |
| China's internet regulator orders an AI app offline over unlabeled output China's Cyberspace Administration (CAC) ordered a WeChat mini-program be taken offline for producing text and images without labeling them as AI output. The case is one of 10 in a bulletin the agency published. The Sichuan technology company behind it added neither the visible labels nor the embedded provider and content identifiers the rules require. Also listed is a Jiangsu company warned over two websites run as relay stations. The sites called several large-model providers' interfaces and offered dialogue and question-answering services without a required security assessment. CAC separately fined a Shanghai technology company for sending users' registration details and host configuration data to data centers abroad without the cross-border security review the law requires. |
|
| Ant Group open-sources a safety check that runs while a model is generating Ant Group's AI Security Lab released SingProbe, a guardrail that works inside a model's generation process instead of screening the finished answer. China's content labeling and model safety rules put a check on every generated response, so the overhead a guardrail adds per response decides whether the compliance layer survives agent-scale traffic. Tested on the lab's Ling-3.0-flash model in production, SingProbe added under 0.5% extra overhead at the decode stage, the point where a model produces tokens one at a time. SingProbe-Med, a version that monitors medical risk during generation, corrected 25.03% of answers the baseline model had originally gotten wrong on the lab's own evaluation. The guardrail has been adapted to 29 mainstream open-source models, among them the Ling-3.0 series, GLM-5.2 and DeepSeek V4, and Ant open-sourced the code, model and benchmark together. |
| AI startup StepFun tops independent voice leaderboards with a new speech model family StepFun released the StepAudio 3 family, now live on its open platform, covering conversation, transcription, synthesis, audio generation and song writing. Speech is the layer through which most people will meet an AI agent, and it rides on hardware China already makes at scale, from handsets to cars. The conversational model took first place on Artificial Analysis's Conversational Dynamics leaderboard with a composite score of 98.9%, and first again on the same firm's Speech Reasoning board at 99.7% accuracy. The transcription model tied for first on that firm's nonstreaming speech recognition board with a 1.7% word error rate, or fewer than two mistakes in every hundred words. StepAudio3Realtime supports full-duplex dialogue, meaning both sides can talk at once rather than taking turns, and reads tone, emotion and ambient sound alongside the words. |
| Chinese Academy of Sciences spinoff open-sources a 9-billion-parameter model for spatial tasks Zidong Taichu, incubated by the Chinese Academy of Sciences' Institute of Automation, open-sourced ZDTaichu5.0-9B, a multimodal model of about 9 billion parameters built for the physical world. Spatial reasoning is the gap between a capable language model and a machine that can act, and open weights at this size put it within reach of anyone who downloads them. The model takes single images, multiple images, long video and visual input at arbitrary resolution, and targets spatial perception, cross-viewpoint transformation and embodied task planning in real scenes. Cross-viewpoint transformation means working out how a scene looks from a vantage point the model was never shown, something a robot needs before it reaches around an object. On the company's own comparison across nine spatial understanding benchmarks, ZDTaichu5.0-9B placed first in its size group on eight. |
| Seven Beijing doctoral students train a 7-billion-parameter model and publish the entire recipe Seven doctoral students at Beijing's Zhongguancun Academy trained ZGCM-1, a 7-billion-parameter model, from scratch, and published everything behind it. Export controls restrict accelerators rather than know-how, so a fully documented training run diffuses wherever the documents go. The release covers the training data and per-stage recipes, the weights, training code, intermediate checkpoints and logs, with the report and code on GitHub and the weights and data on Hugging Face. Combining local and global attention and expanding the training context in stages to 256K tokens gave about 3.94 times the throughput of full attention at that length. On general evaluations, the model performs comparably to same-size peers such as Qwen3-8B, and ranks ahead of them overall across a suite of reasoning tests. |
|
| III | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Robotics startup Junao Panshi claims the first brain-inspired world model for robot planning Shanghai Junao Panshi Technology released Cog-WM 1.0 and calls it the world's first brain-inspired cognitive world model, a planning system for robots. Progress in the field has largely tracked how much human demonstration data a team can collect, and the company's pitch is that architecture can substitute for some of that. A world model is the internal predictor a robot uses to anticipate what happens next before it acts. The architecture pairs brain-like neural mechanisms with JEPA, a design that forecasts future states in a compressed internal representation rather than pixel by pixel. On the company's own reported results, Cog-WM Nav 1.0 raised success at finding a named object in an unseen indoor space from 78.5% to 86.9% against a published baseline. It is already deployed on quadruped robots at customer sites for inspection and patrol work. |
| Sensor maker Xense says its robot touch sensors now hold 97% consistency unit to unit Xense Robotics says its multimodal visuo-tactile sensors now hold 97% consistency from one unit to the next, which it describes as a first for the field. Consistency and durability are what decide whether a tactile sensor stays a research part or becomes a component a factory can buy in volume. Visuo-tactile sensors read touch optically, with a camera watching how a soft contact surface deforms. The upgraded parts reached IP68 waterproofing, the sealing rating for sustained immersion, and ran continuously underwater for 10 days. In magnetic testing they kept working near motors and electromagnets, where tactile designs that read capacitance or magnetic change tend to pick up interference. The flexible contact layer is rated at 5 million press cycles, with force and torque curves the company says stayed consistent from the first trigger to the last. |
| Chinese four-legged firefighting robot runs its first live drill, rated for two minutes at 500 C A domestically-developed firefighting robot called Fanghuo Tianlang completed its first live fire scene drill, moving into simulated building fire rubble ahead of crews to survey conditions. The four-legged structure allows the robot to cross post-fire rubble and damaged stairs and right itself if it tips over. It carries a dual-spectrum thermal imaging gimbal, several categories of gas sensors, a 360-degree camera and a life detection module, relaying data and video live to a command platform. With a heat-resistant protective kit, the machine can work for two minutes in a 500 degree Celsius fire, roughly 930 Fahrenheit. Fanghuo Tianlang does not take part in the main firefighting work; its role is to replace people on high-risk front-line survey tasks. |
Dexmal's robot control model tops all four categories of a new Chinese embodied-AI benchmarkDexmal's embodied foundation model DM0.5 now ranks first on all four sub-leaderboards of RoboColiseum, after the August 25th edition's first place on the separate RoboDojo board. The benchmark was initiated by AgiBot with universities, research institutes and robot companies. Its answer to unreliable scores is calibration, reconstructing real environments in simulation with lidar and calibrating object mass, friction and moment of inertia item by item. Validation showed 89.5% consistency between simulation and reality, with per-task success rates differing by under 10%. Training and evaluation sets are fully separated so that replaying recorded trajectories cannot game the board. DM0.5-equipped robots are moving and sorting goods alongside other robots in a large international retailer's warehouse and assisting packaging and scrap recycling on an electronics production line. |
| Richard Sutton's Openmind co-opens a Beijing site where robots learn by trial and error A robot kindergarten built by Tashan Technology and Openmind, the team of Turing Award winner Richard Sutton, opened at Shougang Park in Beijing's Shijingshan district. Demonstration data is the main constraint on embodied AI, and reinforcement learning from real-world experience is the principal route around it. Rather than train on human demonstrations, the site leaves robots in real physical environments to accumulate experience through perception, action, feedback and reflection. It is organized into test, learning and interaction areas, hosting robots from companies including Unitree and Booster Robotics. Sutton said the conventional mode built on human demonstration data has limits, and that robots need continuous learning through trial and error in real environments. |
|
| IV | QUANTUM & FRONTIER SCIENCE | |
| AI drug-design firm MoleculeMind simulates a full enzyme reaction at near-quantum accuracy MoleculeMind's atomistic modeling platform QuantaMind simulated the complete catalytic cycle of a 17,792-atom plastic-degrading enzyme, with results published in the peer-reviewed journal Science Advances. Reactive molecular dynamics, which tracks chemistry as bonds break and form, has been limited to small systems by the cost of the quantum chemistry method behind it. The paper reports pushing near-quantum accuracy to systems of tens of thousands of atoms over tens of nanoseconds. That range is what it takes to watch an enzyme complete a reaction rather than sample a frozen moment of one. Researchers sampled 342 configurations from the reaction trajectory for independent quantum chemistry calculations and found correlation above 0.99 on atomic force components. |
|
| China Railway rolls out a 16.68-meter tunnel boring machine, the largest China has built itself A tunnel boring machine named China Railway No. 1662, with a 16.68-meter excavation diameter, rolled off the line in Zhengzhou, Henan province. It is the largest China has developed on its own. Heavy construction equipment at this diameter is supplied by few firms worldwide, so where it was built is the point of the announcement. The machine was built jointly by China Railway Tunnel Group, China Railway No. 4 Engineering Group and China Railway Engineering Equipment Group. It is a slurry-balance design, which holds the tunnel face steady with pressurized slurry and is used for soft ground and underwater crossings. Its cutterhead operates at atmospheric pressure, so crews can change cutting tools without pressurizing even where the working face sits at 8 bars, about eight times sea-level pressure. |
|
|
|