|
|
|
|
Bottom Line Up Front CXMT supplied Xiaomi's LPDDR6 phone memory for the Xring O3 processor released yesterday, the first flagship handset chip built for the new low-power memory standard. GalaxySpace delivered an in-orbit satellite to Thailand, launching Lingzhi-09 today from Taiyuan along with two radar satellites and supplying the ground system and operator training to Thailand's space agency. ByteDance released Doubao Work, an office agent that operates a virtual desktop and draws on a company's own Feishu messages, documents and meetings. |
| Chinese memory maker CXMT supplies Xiaomi's first phone chip on the LPDDR6 standard Industry sources name ChangXin Memory Technologies (CXMT) as the memory supplier behind Xiaomi's Xring O3 phone processor, released yesterday, Sina Tech reported. The O3 is the first flagship phone processor built to run LPDDR6, the newest low-power memory standard for handsets. The move challenges Samsung, SK Hynix and Micron, who dominated almost every prior generation. The LPDDR6 line with O3 is scheduled to debut in the Xiaomi 18 Fold, with the CXMT part stacking the memory directly above the processor. The new standard moves to a 24-bit dual sub-channel layout from LPDDR5X's 16-bit channel, which gives more flexibility in scheduling data. |
| GalaxySpace hands a Thai government agency a satellite it built, launched and delivered in orbit A Long March 6C rocket lifted off from the Taiyuan Satellite Launch Center today carrying seven satellites. Among them was Lingzhi-09, a cubesat built by Chinese satellite manufacture GalaxySpace, that will collect remote sensing data on land, agriculture and ecology for Thailand's Geo-Informatics and Space Technology Development Agency. TMTPost said it is the first time a Chinese commercial satellite company has handled whole-satellite development, launch and in-orbit delivery for a Southeast Asian country. The company supplied the ground system and operator training alongside the spacecraft. GalaxySpace also built two synthetic aperture radar satellites on the same launch, ZhongKe-14 and ZhongKe-15, that carry a radar that builds images from returns rather than light. Both receive, compute and analyze on board, so images can be processed within minutes for flood and earthquake response. |
| ByteDance ships an office agent that works on a virtual desktop and reads company data ByteDance released Doubao Work today, an agent that breaks an office job into steps and carries it out by driving a computer and browser on a virtual desktop. Connected to an account on Feishu, ByteDance's workplace messaging and document suite, it draws on a company's messages, documents and meetings, which is what lets it attempt weekly reports, proposals and business analysis. The launch follows a reorganization yesterday that folded the TRAE and Coze teams into the Doubao organization, with their product and operations staff reporting to Doubao product head Zhao Qi. GeekPark attributes the urgency to Tencent's WorkBuddy, which Tencent called China's most widely used productivity AI agent service in its first-quarter results, at roughly one million daily active users. |
| Startup Dexmal open-sources the robot control model now leading the RoboDojo board The Chinese startup Dexmal's DM0.5 took first place on RoboDojo, a leaderboard for models that drive physical robots, QbitAI reported. Dexmal has published the robot control model's weights, its training framework and downstream task workflows on GitHub and Hugging Face. DM0.5 averaged a 19.34% task success rate, against the 8.80% the board's leaders averaged in simulation as recently as July. The benchmark is run by the University of Hong Kong's Multimedia Laboratory with UC Berkeley, Tsinghua University and other institutions. Its widest margin came on memory, where the model natively holds up to 60 seconds of its own prior action sequences and can review what it has already done. |
|
| China's state-asset commission tells central state firms to close 6G theory and component gaps SASAC, the State Council commission that supervises China's centrally owned state enterprises, held a promotion conference on 6G and released a set of scenario tasks for those firms to take up. Its instructions name the shortfalls directly: original innovation in basic theory, breakthroughs in key underlying technologies and components, and a secure, stable supply chain for the mobile network generation after 5G. The commission also called for faster international cooperation and for raising China's international standing, while pushing the technology into the real economy toward commercial use. SASAC figures put central enterprise investment in strategic emerging industries at 2.5 trillion yuan ($350 billion) in 2025, or 41.8% of their total investment. Emerging industries' share of those firms' revenue is up more than 10 percentage points since 2022. |
| MIIT opens comment on the first national standards plans for humanoid robots and brain-computer interfaces MIIT, the ministry that sets China's industrial policy, opened public comment on draft guidelines for national standards systems covering humanoid robots and brain-computer interfaces, with responses due Sept. 23. It is the second standards consultation in two editions, after the automotive cryptography draft covered yesterday. No humanoid robot standard has yet been formally issued in China or internationally, the robot draft states, so the terminology, testing and safety rules it plans would be the first on the books. That plan targets at least 100 key standards by 2028 across capability testing methods, key technologies, applications and safety governance, with more than 200 companies implementing them. The brain-computer interface plan, covering systems that read neural signals and pass them to a computer, targets more than 40 domestic standards by 2028 and a largely complete system by 2030. |
|
| III | AI & FOUNDATION MODELS | |
| Pediatric AI trained on hospital reasoning records outscores general models on Beijing cases Xiao'erfang Health Technology's AI pediatrician reached 89% diagnostic accuracy on real cases from Beijing Children's Hospital, against 68.1% for the best general-purpose model on the same cases, the company reports. It attributes most of the general models' errors to missing expert clinical reasoning rather than thin knowledge, which is the gap its training data is built to close. The system is built on 150,000 records made with the hospital since 2023, each documenting how an expert reasoned toward a diagnosis, including which to rule out. It follows the hypothetico-deductive method taught in medical schools, forming a candidate diagnosis and then testing it against evidence, and checks each step against source documents. It also asks follow-up questions like a doctor. It is in use at hospitals in Beijing, Shaanxi and Jiangsu, starting at Weinan Maternal and Child Health Hospital. |
| Chinese Academy of Sciences spinoff gives its research agent an engine for whole projects Zidong Taichu, a company incubated by the Chinese Academy of Sciences' Institute of Automation, upgraded its research agent ScienceClaw with an engine called AutoProject that takes on a whole research project. Rather than running isolated tasks, the engine turns a loosely stated goal into a network of subtasks, monitors the experiments, reruns the failures and consolidates what comes back. A subsystem called EviGraph builds an evidence graph across those subtasks, testing whether conclusions actually follow from the experimental data. On the ARCBenchML evaluation, the company reports EviGraph scoring 0.865 overall against a best baseline of 0.596, with its largest gain on a measure of how closely conclusions match experimental facts. Worked examples show the engine on a steel-strip surface-defect project and on an object detection workflow split into data processing, modeling and iterative optimization. |
| SenseTime open-sources the full 8-billion-parameter version of its image generation and editing model SenseTime open-sourced the full version of SenseNova U1.5 Lite, an 8-billion-parameter model that generates and edits images in one system, three weeks after the preview. For the full release the company trained separate specialist models for text rendering, aesthetic quality and editing, then folded their behavior back into the single model. SenseTime says the result holds its own against GPT-Image-2, a closed commercial model, on complex typesetting and precise editing at that size. It accepts single instructions of 3,000 to 4,000 characters covering subject, count, position, text and layout at once, and outputs natively at 4K. Editing is addressed by region: a user boxes part of an image and the model changes only that area. Weights and code are posted on GitHub and Hugging Face, with a hosted demo open for testing. |
| Thomson Reuters builds its first in-house model on Alibaba's open-source Qwen to cut Claude use Thomson Reuters adapted Alibaba's open-source Qwen3.5 into its first in-house model, to cut reliance on Anthropic's Claude and long-run costs, PingWest reports. The model, called Thomson, launched yesterday and is deployed first in Tabular Analysis, a structured document review feature in CoCounsel Legal, the company's AI assistant for lawyers. Thomson Reuters puts the training bill at $40 million, against the billions frontier labs spend, and says Thomson will gradually take over work Claude has been doing. Chief technology officer Joel Hron said CoCounsel still relies mostly on Claude; the company says it remains multi-model by design. |
|
| Xiaomi shows a prototype running a 120-billion-parameter model on its own Xring chips A day after launching three in-house Xring chips, Xiaomi founder Lei Jun detailed two prototypes built around them: a phone-sized terminal for on-device AI and a desktop box called the AI Cube. The phone prototype drops the camera module and adds active air cooling, trading hardware a phone needs for thermal headroom. Pairing the Xring O3 with the newer O100, it reached 295 tokens per second on Xiaomi's on-device MiMo model, a measure of how fast the model produces text. The AI Cube combines the O3, O100 and D100 in one body, with 80GB of memory shared across the three chips, enough to hold a 120-billion-parameter model that would otherwise run in a data center. The O3 is scheduled to debut in the Xiaomi 18 Fold; the O100 and D100 are planned for commercial use in 2027. |
|
| V | ROBOTICS & AUTONOMOUS SYSTEMS | |
| MIIT projects 100,000 China-built humanoid robots in 2026, five times 2025 output MIIT, the ministry that sets China's industrial policy, projects national production of complete humanoid robots above 100,000 units in 2026, against 20,000 in 2025, TMTPost reported from the World Robot Conference in Beijing. Industry figures put Chinese manufacturers at more than 90% of global humanoid shipments. The enabling number is cost: average costs for domestically made core components, chiefly reducers, servo motors and sensors, are about 30% below a year earlier, by industry estimates. What is unsettled is engineering rather than invention, with gait control trained in simulation degrading on uneven or oil-covered factory floors and fine manipulation requiring hands with more than 10 degrees of freedom plus force and tactile sensors. A bipedal robot consumes three to five times the energy of a wheeled one, and most current products cannot work an eight-hour factory shift. |
|
|
|