--- date: 2026-09-08 subject: "Kirin stacks two logic layers | Up to 99% duty on a Japanese chip gas | XPeng's humanoid line" --- **Huawei stacked two logic layers** in its 2026 Kirin design, with chief semiconductor scientist Liao Heng saying the dies are joined by roughly 50 million vertical connections, an approach that recovers performance through packaging rather than smaller transistors. **China imposed provisional anti-dumping duties on Japanese dichlorosilane**, requiring importers to post security deposits from today on a chipmaking gas that also feeds Chinese fabrication lines, after preliminarily finding dumping margins of 80.8% to 99.2%. **China's Supreme People's Court set AI liability rules**, telling lower courts they may clear open-source AI developers of liability for a user's infringement when the code was supplied free of charge and the developer publicly explained the model's functions and safety risks. # Top Stories - **Huawei's new phone chip stacks two logic layers with about 50 million vertical connections** — Huawei chief semiconductor scientist Liao Heng said the company's 2026 Kirin design forms about 50 million vertical connections between two stacked dies, of which 5 million to 10 million carry signals. Recovering performance by stacking logic instead of shrinking transistors shifts the question from which manufacturing generation Chinese foundries can reach to how much a designer can extract from the ones already available. Conventional 3D packaging places tens of thousands to a few hundred thousand connections between two dies, Liao said. The part is the Kirin 9050 Pro, which Huawei calls its first logic-folded chip and debuted in the Mate XT 2 tri-fold phone. Chu Junhao, a semiconductor physicist at the Chinese Academy of Sciences, said work across materials, thermal management and packaging still has to follow. Huawei last unveiled a new flagship Kirin part at a launch event with the Mate 40 in 2020. [ITHome](https://www.ithome.com/0/999/716.htm) - **China imposes provisional anti-dumping duties on a Japanese chipmaking gas after finding margins up to 99.2%** — China's Ministry of Commerce (MOFCOM), which runs the country's trade remedy investigations, preliminarily found that Japanese dichlorosilane is being dumped at margins of 80.8% to 99.2%. From today importers must post security deposits at the rate the ruling sets for each company. The same gas feeds Chinese fabrication lines as well as Japanese exporters' order books, so the duty raises the cost of a chipmaking input inside China. Dichlorosilane is a silicon-bearing gas used for logic, memory and analog parts in thin-film deposition, the step that lays down the ultra-thin layers a chip is built from. MOFCOM opened the case on Jan. 7 at the request of Chinese producers of the same gas, the domestic industry it found had been injured. A ministry spokesperson said China uses trade remedies prudently and with restraint, and that the investigation continues toward a final ruling. [China Military Online](http://www.81.cn/ss_208539/16484218.html) [govt] - **China's top court sets liability rules for AI deepfakes and open-source models** — China's Supreme People's Court, which writes the guidance lower courts work from, has told them they may clear an open-source AI developer of liability for a downstream user's infringement. That exemption turns on disclosure rather than license choice, so a Chinese model release now carries a legal reason to document what the model does and where it fails. It applies where the developer supplied the code free of charge and publicly explained its functions and safety risks. The Opinion on Hearing AI-Related Dispute Cases also covers AI face-swapping and voice cloning, AI resurrection of deceased people and online doxxing. Attorney You Yunting, writing in Huxiu, measured four flagship Chinese releases against that test and found that none explains what the model does or addresses hallucination or deepfake risk. DeepSeek-R1 ships under the standard MIT license and Alibaba's Qwen3.8-27B under Apache-2.0, both general-purpose open-source licenses that end in a disclaimer. [Caixin](https://china.caixin.com/2026-09-08/102482710.html) [Huxiu](https://www.huxiu.com/article/4889479.html) - **XPeng opens what it calls the world's first automated humanoid robot production line** — XPeng switched on what it calls the world's first automated production line for general-purpose humanoid robots, with its Iron robot completing final assembly and walking off the line under its own power. XPeng designed the line in-house and puts core-process automation above 80%, which moves the constraint on humanoid robots from whether one can be built to how many can be built the same way. The Iron robot runs three of XPeng's in-house Turing AI chips, which the company puts at up to 2,250 TOPS of effective compute, enough to run its physical AI foundation model on the robot itself. TOPS counts trillions of operations per second, the standard measure of AI accelerator throughput. Chairman and chief executive He Xiaopeng said the robot production lines "were created from scratch with no precedent to follow." XPeng plans volume production of Iron by the end of 2026, with launch and deliveries in China and overseas markets in 2027. [PingWest](https://www.pingwest.com/w/317190) [CnEVPost](https://cnevpost.com/2026/09/08/xpeng-opens-iron-humanoid-robot-production-line/) - **Chip designer Axera launches 5nm driver-assistance chips rated up to 720 TOPS** — Axera has launched two 5nm driver-assistance chips, the flagship M97 rated at 720 TOPS of equivalent AI compute and the mainstream M95 at 360 TOPS. TOPS counts trillions of operations per second, the usual yardstick for AI accelerator throughput. The M97's in-house image signal processor handles 16 cameras, and both parts carry an on-die safety island held to ASIL-D, the most demanding automotive safety level, inside an overall ASIL-B rating. Axera says both chips run the newer driving models that turn camera input and instructions directly into vehicle commands, at performance it puts significantly above industry level. The company positions the M97 as its flagship tier and the M95 as its mainstream tier, both on the same 5nm process and both supporting dual-chip redundant configurations. [ITHome](https://www.ithome.com/0/999/803.htm) # Policy & Regulation - **Chinese court convicts six who stripped an AI model's filters to sell obscene images** — A Zhejiang court has convicted six people of producing obscene material for profit after they stripped the safety limits from a generative AI model and charged users for the output. The conviction places criminal responsibility on the developers who altered the model rather than on the users whose prompts produced the images. The jailbreak worked by repeatedly rewriting system prompts, setting character cards and building dialogue scaffolds until the model's built-in pornography limits gave way. The service launched on an overseas platform in January 2025 and generated more than 370,000 static images, with user top-ups of about $24,000. Judge Zhu Zhanglin of the Yuyao People's Court said a developer who deliberately alters a model's safety mechanisms remains the producer under criminal law, even though users supplied the prompts. [ITHome](https://www.ithome.com/0/999/811.htm) - **China's industry ministry makes 6G base stations and handsets five-year plan targets** — China's Ministry of Industry and Information Technology (MIIT), which licenses telecom operators and sets industrial policy for the sector, issued its Information and Communications Industry Development 15th Five-Year Plan. The plan tells the industry to start 6G commercialization at an appropriate time, wording that sets no date and leaves the schedule to the technology trials it also orders continued. Domestic development of 6G base stations, core network equipment and 6G smartphones is directed under a stated goal of strengthening supply chain resilience. OpenHarmony, the open-source operating system project behind Huawei's HarmonyOS, is named as a base for terminal operating systems supporting agent communication and environmental sensing. MIIT routes 6G technology trials through the IMT-2030 (6G) Promotion Group, the state-convened body coordinating China's 6G research, tasking it with a test-verification platform, and directs work on satellite and terrestrial network integration. [Global Times](https://china.huanqiu.com/article/4T7tTukJZHW) [state media] # AI & Foundation Models - **DeepSeek puts a cheaper multimodal model into closed beta on a new architecture** — DeepSeek told its official developer group that an interim V4.1 Flash build has entered closed beta, on a new model architecture with native multimodal support. Billing matches the existing DeepSeek V4 Flash endpoint, so the build arrives as a drop-in substitute at an unchanged price for developers already there. DeepSeek's beta feedback survey asks users whether the interim build could fully replace the online V4 Pro, putting a cheaper model forward as a substitute for a more expensive one in production. Native multimodal support means the model takes images and text through one pathway rather than bolting a vision model onto a text model. Beta users keep the same API address and switch the model name to a string carrying a Sept. 10 expiry, with a cap of 20 concurrent requests per account. [ITHome](https://www.ithome.com/0/999/795.htm) - **ModelBest open-sources a small model built to run agents on devices** — ModelBest and the OpenBMB open-source community have released MiniCPM5-2B, an on-device language foundation model that scores 20 on the Agentic Index against same-class models all sitting below 10. If a model this small can hold an agent loop, agent work can run on phones and cars instead of the cloud models it leans on today. GeekPark said cloud inference cost and data security are the constraints on deployment. An agentic benchmark measures whether a model can carry out multi-step tasks using tools, not answer a single question. On-device models run on the phone, car or appliance itself rather than calling a data center. ModelBest released the full training recipe, the reinforcement-learning framework and the datasets rather than the weights alone, letting other teams reproduce the training instead of only running the result. [GeekPark](https://geekpark.net/news/369999) - **Shenzhen startup Atomelody says its molecular-structure model beats AlphaFold 3** — Atomelody, founded four months ago in Shenzhen, released Melo-1 Preview and says it exceeds AlphaFold 3 on most evaluation metrics for peptides, small molecules and antigen-antibody complexes. A Chinese commercial model at that level matters as much for access as for accuracy, because AlphaFold 3's weights remain closed to commercial organizations under Google DeepMind's usage terms. AlphaFold 3 is Google DeepMind's structure-prediction model, and structure prediction is the foundation layer for AI driven drug design. Atomelody says it built physical and chemical constraints into the architecture so that high-scoring outputs stay physically stable, avoiding models that place atoms in collision or produce bond lengths that cannot exist. The company says its core team previously used AI to crack more than 10 GPCR targets, the cell-surface proteins that rank among the most common drug targets. [BAAI](https://hub.baai.ac.cn/view/57771) [state-affiliated] # Chips & Semiconductors - **Memory maker CXMT says its DRAM now matches international mainstream suppliers** — CXMT said, after reports of a cooperation relationship with Apple, that its products can already compete with international mainstream manufacturers on performance, quality and supply stability. That is a wider claim than the LPDDR6 design win in the [August 25th](https://chinatechdaily.org/archive/china-daily/2026-08-25/) edition. Whether a U.S. flagship device maker qualifies Chinese DRAM is the practical measure of how far memory now sits outside the export control perimeter. CXMT, or ChangXin Memory Technologies, makes dynamic random-access memory, the working memory a processor reads and writes as it runs. Rather than address the Apple reports directly, the company said it is exploring cooperation opportunities with quality global customers with an open attitude. Sina Tech reported that CXMT already has deep cooperation underway with Alibaba Cloud, ByteDance and Tencent, alongside handset brands including Honor, OPPO and vivo. [Sina Tech](https://finance.sina.com.cn/jjxw/2026-09-07/doc-iniqyzxs9627478.shtml) [state-affiliated] - **Automaker Huajing extends delivery times over a cockpit chip shortage** — Huajing Auto has extended delivery cycles for the Huajing S and apologized, blaming a phased supply shortage of the in-vehicle infotainment chip paired with the car's HarmonyOS cockpit system. A carmaker that says its plants are fully ready but cannot get one cockpit chip puts the constraint in the chip supply chain rather than in the factory. A cockpit chip runs the dashboard and infotainment software, and is separate from the silicon that handles driving functions. The company said its manufacturing capacity is fully ready, while its own forecasting and contingency planning for industry-wide chip supply challenges were insufficient. Order volumes for the Huajing S have climbed steadily since launch, compounding the delay, and Huajing says it is expanding chip sourcing channels with core partners and pursuing diversified supply. [Gasgoo](https://auto.gasgoo.com/news/202609/7I70471347C501.shtml) # Robotics & Autonomous Systems - **Huawei says its assisted-driving and cockpit systems are now in 2 million vehicles** — Huawei's Intelligent Automotive Solution unit said Qiankun assisted-driving installations and HarmonyOS cockpit installations have each reached 2 million. Installation counts matter beyond the sales figure, because road data at that scale is the input the next generation of the driving stack trains on. The unit supplies driving and cockpit systems to other carmakers rather than building a Huawei-branded car. Cumulative mileage driven under Qiankun ADS, Huawei's Advanced Driving System, has passed 14.9 billion km. Jin Yuzhi, the unit's chief executive, said in July that Huawei will invest 70 to 80 billion yuan (about $10 billion to $12 billion) in Qiankun compute research and development over the next five years. [EV Industry News](https://www.evchanye.com/2026/0908/52290.shtml) - **Robot maker Yuequan puts a wheeled humanoid to work on live power lines** — A wheeled humanoid robot has stripped insulation from energized electrical cables for State Grid in Hefei, completing the full workflow without the power being cut. Live-line maintenance is among the more dangerous jobs in a power network, and the alternative to doing it energized is cutting supply to the customers downstream. The machine is the Bowen W-Bot 2.0 from Yuequan Bionics, mounted on an insulated bucket truck so a crew can raise it into position. State Grid, the state-owned operator that runs most of China's electricity network, is the customer, and version 2.0 ran live-line testing at the 2026 World Robot Conference before this Hefei deployment. Balance is the harder problem, since a single arm lifting 10 kilograms off-axis will tip an unsecured platform, so the robot computes its center-of-mass offset and redistributes load across the joints. [EV Industry Technology](https://www.evchanye.com/2026/0908/52266.shtml) - **D-Robotics humanoid clears stairs and gaps on a controller trained only in simulation** — A humanoid running D-Robotics' new walking controller cleared stepping stones, stairs and gaps at a mean 93.3% success rate, across 10 trials on each obstacle. Removing hand-tuned stages matters for cost, since each one is engineering labor redone for every new robot and every new terrain. The controller, WM-LOCO, was trained purely in simulation, and the same weights went onto a Unitree G1 with no retuning. Its claimed novelty is bringing a predictive world model, a learned internal simulator of how an action changes what the robot senses, into walking over complex terrain. D-Robotics says the approach needs no hand-labeled footstep targets, no pre-planned foot trajectories and no teacher-student distillation. On the gap tests the robot crossed 0.8 meters in one motion, and the team stopped widening the gap for safety rather than because the controller failed. [EV Industry Technology](https://www.evchanye.com/2026/0908/52265.shtml) # Military Watch - **China and Singapore move their naval exercise from port to sea off Zhanjiang** — Singapore's frigate RSS Steadfast and the Chinese navy hull Dali will leave Zhanjiang for the sea phase of the China-Singapore Maritime Cooperation-2026 exercise (covered on [September 4th](https://chinatechdaily.org/archive/china-daily/2026-09-04/)), moving into the sea and air space nearby. Submarine rescue is a mission where a disabled boat depends on whichever navy can physically reach it, so the exchanges build relationships a navy falls back on in a subsurface emergency. Steadfast arrived at a naval port in Zhanjiang, Guangdong yesterday, and officers from both sides toured both ships before the harbor phase closed. The two navies held professional exchanges on three subjects: diving operations, submarine search-and-rescue and underwater medicine. Specialists from the Southern Theater Command Navy, the fleet responsible for the South China Sea, took part alongside the Naval Medical University. [Global Times](https://mil.huanqiu.com/article/4T7vrVmNje5) [state media] - **China closes part of the South China Sea for four days of live-fire gunnery** — The Zhanjiang Maritime Safety Administration has closed a designated area of the South China Sea for live-fire gunnery, in a navigation warning published on the national maritime safety agency's website. A navigation warning is how a maritime authority reserves sea space from commercial traffic, so the closure window doubles as the published schedule of the firing. The firing runs from today through Sept. 11, each day from 8 a.m. to 6 p.m. Entry into the area is prohibited for the duration. [Global Times](https://mil.huanqiu.com/article/4T7GwvKREy7) [state media] - **Rocket maker LandSpace rolls out a launcher for China's satellite internet network** — LandSpace posted photos yesterday of its Zhuque-2E rocket being rolled out, pointing to an imminent launch of networking satellites for the Qianfan broadband constellation. Filling the planned slots depends on launch capacity rather than satellite manufacturing, since Qianfan and the GW constellation together call for more than 28,000 satellites. China now has two parallel systems for recovering and reusing rocket first stages, Tebon Securities said, after the Zhuque-3 made the country's first land recovery of an orbital first stage on landing legs. That sits alongside the Long March 10B's sea-based net recovery, and reusing a first stage, the costliest part of a rocket, is what brought launch prices down elsewhere. Domestic rocket launches grew 29.3% year-on-year from January through July 2026 while satellites placed into orbit grew 66.9%, per Haitong International Securities, meaning more satellites rode up per flight. [Jiemian](https://www.jiemian.com/article/15069067.html) # Strategic Context - **China opens a fourth AI training workshop for developing-country officials** — China's Ministry of Foreign Affairs opened its fourth AI capacity-building workshop at Peking University yesterday, drawing nearly 40 senior officials from founding members of the World Artificial Intelligence Cooperation Organization (WAICO). The scaffolding is the substance here, since the training pipelines and the organization are where AI governance norms for developing countries get set. The session implements a pledge President Xi Jinping made in July in Shanghai to provide developing countries with 5,000 AI training slots over five years. Foreign Ministry spokesperson Mao Ning said WAICO is meant to help Global South countries find AI development paths suited to their own needs. Renan Mendes Gaya Lopes Dos Santos, who directs public data infrastructure at Brazil's management and innovation ministry, called China's sharing of open-source AI models and training "a model for us." The session runs through Sept. 12. [Yicai](https://www.yicai.com/brief/103356110.html) [state-affiliated] [Global Times](https://world.huanqiu.com/article/4T7tN1uiiFA) [state media] [Bastille Post](https://www.bastillepost.com/global/article/6136904-4th-ai-capacity-building-workshop-opens-in-beijing)