← China Tech Daily
About Editorial Standards Corrections Privacy Policy
Subscribe →
China Tech · Daily chinatechdaily.org
OCTOBER 14, 2026
Your morning brief on Chinese technology. Published by the AI Policy Institute
Bottom Line Up Front

ByteDance's AI research unit Seed creates safety department to guard against chemical, biological, radiological and nuclear misuse of the models that power its Doubao chatbot, which had 382 million active users in June. WeChat, DeepSeek and 149 other apps refuse access to Doubao Phone Assistant, the ByteDance smartphone agent that completes tasks inside other apps by reading their screens and simulating taps. Startup Landau Flint shows EUV light source prototype, the component that generates the beam advanced chipmaking machines print with, at a Shanghai industry fair, where booth staff put mass production 3 to 5 years away.

ITOP STORIES
 
ByteDance's AI research unit sets up department to guard against chemical, biological and nuclear misuse
ByteDance Seed, the company's large-model research unit, created a Model-AlignmentAlignmentThe problem of making an AI system reliably pursue the goals its developers and users intend, and the research field devoted to it. Misalignment covers everything from a chatbot flattering users to a capable system deceiving or resisting its operators. & Preparedness department. Seed's models power Doubao, ByteDance's chatbot with 382 million active users in June, so a dedicated unit now evaluates AI that reaches a mass audience, including for chemical, biological, radiological and nuclear misuse. The department absorbs Seed's Responsible AI team and covers safety evaluation, risk monitoring and alignment training, which teaches a model to behave as its developers intend. Its other stated goal is safer AI agentsAI 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., systems that carry out multistep tasks on their own. Policy guidance that China's internet regulator and other agencies issued in May already says agents must not act beyond their authorized scope.
Read at Wallstreetcn ↗
WeChat, DeepSeek and 149 other apps bar ByteDance's phone AI assistant from operating them
Doubao Phone Assistant, ByteDance's AI agent for smartphones, published a list of 151 apps that have refused to let it read their screens and simulate taps, including WeChat, Taobao and DeepSeek. Completing tasks inside other apps is the assistant's main function, so the refusals keep it out of many of the messaging, shopping and payment apps where everyday tasks happen. Tencent-linked apps account for 65 of the refusals and Alibaba's for 23, while independent refusers include Kimi, Pinduoduo and major banks. With its public notice period over, the assistant will gradually open automation to apps that have not refused, on the ZTE Nubia phones that ship it.
Read at Wallstreetcn ↗
Chinese startup Landau Flint shows prototype light source for extreme ultraviolet chipmaking machines
Landau Flint, founded in 2024 on research from a Tsinghua University plasma lab, showed a prototype extreme ultraviolet (EUV) light source at the China International Industry Fair in Shanghai. EUV machines print the finest features on advanced chips, and the light source generates the beam they print with, so any Chinese-built EUV scanner would need one. Booth staff said 90% of the system is domestically sourced and put mass production 3 to 5 years away, with output power and long-run stability still to be solved. Commercial scanners need hundreds of watts, and the company's website lists its low-power source, now being readied as a product, at more than 15 watts of EUV output.
Read at ITHome ↗
Chinese AI chip makers show full server racks, and one begins delivering training clusters
Iluvatar CoreX, a Chinese designer of graphics processors for AI, debuted a rack linking 144 of its chips at the China International Industry Fair in Shanghai, where MetaX and Enflame also showed rack-scale systems. Large AI models are trained on many linked chips, so complete racks put these Chinese accelerators in the form buyers actually train on. Iluvatar's own calculations put the rack's cost of processing text about 40% below mainstream international systems, a figure Jiemian says still needs testing on real workloads. Another domestic chip designer, Dongfang Suanxin, says its first large training clusters have entered delivery and that limited manufacturing capacity is capping output. It plans to deliver a thousand-card cluster by year-end.
Read at Jiemian News ↗
IIPOLICY & REGULATION
 
Hunan plan encourages government bodies and state firms to favor Huawei-derived operating system
Hunan province issued a plan encouraging its party and government bodies and state-owned enterprises to give priority, where secure and technically feasible, to OpenHarmony, the open-source operating system derived from Huawei's HarmonyOS. Because the preference reaches government offices and state companies, it could steer which operating systems one province's public sector adopts. The plan assigns Hunan Iron and Steel Group and heavy-machinery makers Sany Group and Zoomlion to build industry-grade versions with Harmony ecosystem partners. Its 2028 targets include 100 demonstration scenarios, 200 applications and 50,000 developers in the province.
Read at ITHome ↗
IIIAI & FOUNDATION MODELS
 
Hunan party newspaper and 360 launch service to shape what AI chatbots say
Hunan Daily, the provincial Communist Party newspaper, and internet security company 360 Group launched Zhimou Trusted GEO, a "generative engine optimization" service that tailors content so AI chatbots draw on it. By 360 President Ye Jian's account, the service also monitors and corrects what large AI models output, so the paper's content would reach users through chatbot answers as well as print and search. Ye said the aim is for authoritative content to reach users accurately, completely and positively, and that the service runs on 360's own models without a third-party base model. A deputy head of Hunan's provincial propaganda department attended the launch, where a first batch of government and corporate clients signed on.
Read at PingWest ↗
AI startup ModelBest releases open-source 35-billion-parameter model that runs offline on phones
ModelBest released MiniCPM-V 4.7, an 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. that reads images and video as well as text, in a 35-billion-parameter version it says runs on phones and tablets in about 3GB of memory. Running a model that size offline lets capable image-and-text AI work on a device without sending data to a cloud server, and the open weights let any developer build on it. The model is a mixture-of-experts design that activates only a small fraction of its parameters at each step, and ModelBest's scheduler preloads the ones needed next so the device does not stall. QbitAI reported more than 13 tokens, or word fragments, per second offline on an iPhone 17 Pro Max.
Read at PingWest ↗·Read at QbitAI ↗
IVCHIPS & SEMICONDUCTORS
 
Specialized AI chips outshipped graphics processors in China's data centers for first time
AI chips other than graphics processors, such as Huawei's Ascend line, outshipped graphics processors in China's data centers for the first time in the first half of 2026, per market research firm IDC. If a Huxiu column's estimates hold, how many more domestic AI chips reach those data centers depends on what SMIC, China's largest contract chipmaker, can make, since demand already exceeds supply. One institutional estimate would leave roughly 40% of 2026 domestic AI chip demand unmet by SMIC's advanced N+2 process. Nvidia still led individual vendors with a 34% share, ahead of Huawei's 28%.
Read at Huxiu ↗
Memory startup MemSilicon sends first high-density resistive memory chip to manufacturing
MemSilicon, a Chinese memory chip startup, has sent its first resistive memory (RRAM) chip, the 1-gigabit MountTai 2026, to manufacturing and is now validating the parts. AI chips sit idle waiting on data from memory, and RRAM, which stores data by switching a material's resistance and keeps it without power, is one way to put denser storage beside the processor. The chip shares one transistor among four memory cells built into its upper wiring layers, so it can be made on a standard 28nm production line. MemSilicon raised more than 100 million yuan ($15 million) in January in a round led by China Development Bank's venture arm and Lenovo Capital.
Read at ITHome ↗
VROBOTICS & AUTONOMOUS SYSTEMS
 
Electric-car maker XPeng offers its vehicle platforms, driving software and chips to other companies
XPeng Chairman He Xiaopeng announced at the Paris Motor Show that the Chinese electric-car maker will open technology cooperation to partners worldwide, offering vehicle platforms, modules and a development toolchain. Offering these to outside firms would put XPeng's driver assistance software and in-house Turing AI chips into vehicles other companies build. The modules cover driver assistance, an AI computing box, chips, batteries and electric drive, and XPeng is also offering packages for its humanoid robots, robotaxis and flying cars. He said the robotaxi business is in in-depth talks with several leading European mobility platforms, with progress expected to be made public soon.
Read at ITHome ↗

Thanks for reading. See you tomorrow.— Daniel

Comments or corrections? hello@theaipi.org

 
Editor-in-Chief:Daniel Colson
Managing Editor:Min Goodman-Cheng
China Policy Fellow:Tyler Liggett
Lead Software Engineer:Christopher Käck
Research Manager:Philip Wieczorek
Operations Specialist:Joanne Chua
Research Assistants:Claude Code + Codex
 

View in browser · LinkedIn

AI Policy Institute · 700 Pennsylvania Avenue SE, Washington, DC 20003

China TechDaily
About Editorial Standards Corrections Privacy Policy
© 2026 AI POLICY INSTITUTE · WASHINGTON D.C.