|
|
|
|
Bottom Line Up Front Alibaba's T-Head released the Zhenwu V900, an AI accelerator that the company says delivers three times the compute of its predecessor and can scale to a single cluster of 500,000 cards. Xiaomi open-sourced its MiMo-V2.6 series in Pro and Flash versions, with the Pro version scoring 46 on the Artificial Analysis composite intelligence index, the highest score for any open-source model. Alibaba said Qwen3.8-Max trained itself, describing 33 effective rounds of iteration in which the model built its own training pipeline, constructed its own training data and located its own defects without human involvement. The development follows previous claims of early recursive self-improvement by Zhipu AI this month. |
| Alibaba chip unit releases an AI accelerator it calls China's most powerful T-Head, Alibaba's in-house chip design unit, released the Zhenwu V900 accelerator today at the Apsara Conference in Hangzhou. U.S. export controls restrict Chinese buyers' access to the highest-end American AI chips, so designing its own puts Alibaba Cloud's computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. plans outside the reach of those rules. The company says the accelerator delivers three times the compute of its previous generation, the Zhenwu M890. A single cluster can scale to 500,000 cards, enough to train and serve frontier-scale models. Chief executive Wu Yongming set a target of more than 20 gigawatts of Alibaba Cloud data center capacity worldwide by 2032. Wu said mid-to-long-term demand for compute far exceeds what Alibaba can supply, with a global shortage across AI data center supply chains limiting how fast its compute base grows. |
Xiaomi open-sources its MiMo-V2.6 models and takes the top open-model score on a third-party indexXiaomi released and open-sourced the MiMo-V2.6 series today in Pro and Flash versions, its first new MiMo model in five months. It is the finished product behind the mid-run training dashboard CTD reported September 17th. Both versions are natively omni-modal, meaning one model handles text, images, audio and video rather than bolting separate components together. MiMo-V2.6-Pro scored 46 on the composite intelligence index published by Artificial Analysis, an evaluation service that blends several reasoning and knowledge tests into one number. That places it above Moonshot AI's Kimi K3 and Alibaba's Qwen3.8-Max, and makes it the strongest open-source modelOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. and the strongest Chinese model on that index. Xiaomi's large-model team attributes the gain to spending more compute on reinforcement learning. |
Alibaba says a Qwen model ran 33 rounds of self-directed training with no human involvementAlibaba said at its Apsara Conference that Qwen3.8-Max built its own training pipeline, constructed its own training data and located its own defects, across 33 effective rounds of iteration with no human involvement. The company calls this recursive self-improvement, a loop in which a model does the work of improving the next version of itself. It follows Zhipu AI's claim of an early form of recursive self-improvement, reported in the September 17th edition. The same loop now reaches the hardware, with Qwen3.8 running front-end design, verification and back-end layout on its own from one real bus-module specification. That run made more than 10,000 calls to electronic design automation tools, the software used to lay out chips, and cut physical-implementation area by 42%, the company reports. Qwen3.8-Max ranked first among agenticAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. models on Artificial Analysis and first in CodeArena front-end coding. |
| Alibaba's T-Head publishes its first server CPU roadmap and dates a fully in-house core T-Head published its Yitian server processor roadmap for the first time, naming the Yitian 720 and the Yitian 730 for the third quarter of 2027. The unit already supplies the AI accelerators Alibaba Cloud runs on, so committing dates for its own host processors takes another imported component out of those racks. Server CPUs are the host processors that manage memory, storage and the accelerators alongside them in a data center rack. The Yitian 730 becomes T-Head's first processor on a microarchitecture it designed entirely itself. T-Head says its single-core performance per gigahertz on the SPECint2017 integer test reaches up to 1.4 times the Yitian 710's. Further out, the Yitian 750 will support ICN, T-Head's own chip-to-chip interconnect, letting its processors connect directly to Zhenwu AI chips. |
| China's weather agency plans to put its AI weather warning system into Central Asia by 2027 The China Meteorological Administration, the national weather service, released a two-year action plan for taking its Mazu early warning system into Central Asian countries. Mazu's forecasts are built from Chinese satellite data and Chinese forecasting models, including AI models. The plan sets out workstreams including co-developing warning products, building shared observation and data systems and training meteorologists in recipient countries. A customized Nepal version reached that country's meteorological bureau for trial use on Aug. 29, days after a mudslide there caused heavy casualties. That build combines observations from the Fengyun weather satellites with the domestic AI forecasting models Fengqing and Fengshun to produce monitoring, forecast and warning products at high resolution. China has pledged to bring Mazu to 30 countries, and the administration says it is running in seven and on cloud trial in more than 50 others. |
|
Zhipu's GLM-5.3-Flash takes the top daily usage slot on the API marketplace OpenRouterZhipu AI's GLM-5.3-Flash took first place on OpenRouter's daily model usage leaderboard, logging 4.36 trillion tokens of calls. That is traffic on the base model behind the faster FlashX tier priced in the September 21st edition. Zhipu is winning that traffic on price and availability rather than on any hardware advantage. OpenRouter is an aggregator that routes developer requests across many models, which makes its daily board a rough census of what Western developers run. Artificial Analysis puts the model at roughly nine cents a task, against 67 cents for OpenAI's GPT-5.6 Sol at a marginally higher intelligence score. |
| Alibaba ships enterprise AI agents carrying an organizational identity and an off switch Alibaba's Qwen Office released a set of enterprise AI agent products at the Apsara Conference, which group vice president Chen Yusen called the industry's first enterprise-grade agent product. Chen said agents used inside companies need clear limits on what they can do and a record of every task they carry out. The digital employees feature gives an agent an organizational identity with a name, a department, a responsible owner and an authorization scope. Administrators can disable an agent, and every execution record traces back to its identity. |
| Database maker OceanBase tops a data agent benchmark running on a Chinese model OceanBase's entry to the Data Agent Benchmark scored 90.62% accuracy, taking first place and clearing 90% for the first time on that leaderboard. That entry, submitted under the internal codename Scout, runs on the domestic GLM-5.2 large language model and beat entries built on GPT, Claude Opus and Claude Fable. UC Berkeley's EPIC Data Lab and Hasura's PromptQL run the benchmark, which tests whether an agent dropped into a scattered data environment can reach a correct answer and show that it is correct. OceanBase says the capabilities will be folded into its DataPilot product for AI driven data analytics. |
|
| III | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Zhejiang sets up a state-backed joint venture to collect robot training data Zhejiang Embodied Intelligence Data Operation Co. has been established with registered capital of 90 million yuan ($13 million), funded by four parties as the province's embodied intelligence data operator. Embodied intelligence pairs an AI system with a physical body such as arms, grippers and sensors, so training it requires recordings of physical tasks rather than web text. Wasu Media holds 55% and supplies access to real production and daily life settings across Zhejiang, with the Hangzhou Gaoxin Shushang holding 33% and handling rights and authorized circulation. Collection runs from first-person video and wrist-mounted cameras up to teleoperation of real machines, with pricing and quality measured against a single unit the company calls the effective hour. |
|
| IV | QUANTUM & FRONTIER SCIENCE | |
| Yang Liwei recaps China's two-rocket plan for landing astronauts on the Moon Yang Liwei, deputy chief designer of China's crewed space program, described the previously published lunar landing plan at a State Council Information Office press conference yesterday. Two rockets will launch separately, one carrying the Mengzhou crew spacecraft and one the Lanyue lander, which will rendezvous and dock in lunar orbit. Astronauts transfer into the lander, descend to the surface, then lift off and re-dock before returning to Earth aboard Mengzhou. Yang said the round trip, lunar-orbit rendezvous and docking, precision landing and liftoff, and the life support system that keeps a crew working on the surface all remain open problems. At 380,000 kilometers from Earth, much of that work has no backup and no rescue option, he said. China plans to land astronauts on the Moon before 2030. |
| Five state enterprises co-organize a quantum competition with a dual-use track Five central state-owned enterprises are co-organizing the first Central SOE quantum applications competition, which opened in Hefei yesterday with entries sorted into tracks covering communications, dual-use technology and energy, among others. The five are AVIC, China's state aviation and defense manufacturer, along with State Grid, China Telecom, China Minmetals and China Merchants Group. They jointly released more than 50 real business-scenario requirements, among them ultra-early detection of partial electrical discharge in high-voltage equipment. Teams must build on a domestic quantum programming framework, which the organizers frame as advancing a self-controlled quantum software toolchain. Winning projects enter a six-month incubation program backed by national-level funds, central-enterprise capital and local state capital. |
|
| China Unicom's design institute releases a second-generation secure phone for government, military and state enterprise users CITC, the design institute of state carrier China Unicom, released Quantum-Secure Phone 2.0 at the carrier's Sept. 9-10 partner conference in Shanghai. The base hardware is a customized version of mainstream domestic flagship phones, so the secure handset build on existing supply chains. The security package pairs an encrypted messaging application with a SIM card that combines a quantum random number chip and a chip implementing guomi, China's national cryptography standards. Version 2.0 runs an offline security risk-control model and an encrypted voice transcription engine on the device itself. CITC says the phone is in scaled deployment across party and government bodies, discipline inspection organs, the military and central state-owned enterprises. |
| Chinese navy trains crewed aircraft and drones as one reconnaissance network A People's Liberation Army Navy unit flew a special-mission aircraft over a designated sea area together with several drones from another unit, all networked onto a shared data link. That division of labor is the maritime surveillance architecture a navy would use to find and track ships at range. The crewed aircraft provided wide-area early warning cueing from high altitude while the drones ran close-in reconnaissance at low altitude. Participants then simulated heavy electromagnetic jamming, which made the radar display flicker and destabilized the video coming back from the drone. A radar operator switched detection frequency bands to characterize the jamming source, while a drone operator used terrain to break line of sight. The drone fell back to a backup data link to keep imagery reaching the command terminal. |
|
| Server maker Inspur doubles its top AI machine to 128 chips in one node Inspur Information launched a super-node AI server called the Metabrain SD200 Ultra at this year's China AI Computing Conference, raising the number of accelerator chips in a single node from 64 to 128. A super-node packs many accelerators into one tightly coupled machine so a model too large for a single server can run without being split across a network. The machine can hold and run Kimi K3, an open-source model of 2.8 trillion parameters, on its own. IDC, a technology market research firm, projected at the same conference that the share of global AI compute demand actually met will fall to 71% by 2027. |
| Automaker Dongfeng builds a shipping-container power station that makes hydrogen from sunlight Dongfeng Motor has built what it calls the industry's first integrated solar-hydrogen-electric energy station, installed in a village in Kalpin county, in Xinjiang's Aksu prefecture. Packaging off-grid hydrogen generation as a shipping container turns it into something that can be sold abroad rather than a site-specific installation. The whole plant fits inside a 20-foot container and combines photovoltaic generation, battery storage, hydrogen production by PEM electrolysis, solid-state hydrogen storage and a fuel cell. PEM electrolysis splits water using electricity across a proton exchange membrane. Dongfeng says the equipment runs between minus 20 and plus 50 degrees Celsius and withstands sandstorms, with a service life of seven to 10 years. It is projected to save the area around 100,000 yuan ($15,000) a year in electricity costs. |
|
|
|