CXMT's denser DRAM passes sample tests | Meituan trains huge AI model on Chinese chips | Apple sets Qwen powered AI for China

Your morning brief on Chinese technology.

Bottom Line Up Front

Chinese chipmaker CXMT reported DRAM trial chip validation for its sixth-generation process, whose 4F² layout gives each memory cell about two-thirds the silicon of today's standard design. Meituan says it trained LongCat-2.5 on Chinese chips, and Leiphone reports the food delivery company's 1.6-trillion-parameter model is the largest training run completed on Chinese hardware to date. Apple and Alibaba set December China launch for Apple Intelligence features that can send iPhone users' requests in Siri and Apple's writing and camera tools to Alibaba's Qwen model. AI firm ModelBest's agent outperformed DeepSeek's chip code with a kernel that ran 9.8% faster on Nvidia's B200, QbitAI reports, and Huawei's Ascend software team said the open-source agent topped three of its kernel performance leaderboards.

ChangXin Memory Technologies (CXMT), a Chinese maker of DRAM, the working memory in computers, said trial chips from its sixth-generation process, which it now says is built on the 4F² cell design it reported progress on in the October 13th edition, have passed validation. Each memory cell in the process's 4F² layout takes about two-thirds the silicon of today's standard design, so a working process would put more memory on every wafer. Sina Tech judges CXMT the first company to report reaching this stage with 4F², a layout Samsung and SK hynix are also developing. CXMT's research spending rose 73% year on year to 11.07 billion yuan ($1.7 billion) in the first three quarters, PingWest reports.

Read at Sina Techstate media ↗·Read at PingWest ↗

Food delivery company Meituan says it trained a trillion-parameter AI model entirely on Chinese chips

Meituan released LongCat-2.5, a 1.6-trillion-parameter model built for long, multistep tasks, and says it trained the model entirely on domestic chips. If that account holds, a model of that size was trained end to end without imported AI chips, in what Leiphone reports is the largest training run completed on Chinese hardware to date. Because each Chinese chip carries little memory, Meituan worked with the chipmakers to keep most data traffic inside supernodes, clusters of chips joined by a fast internal network. In Meituan's own benchmark results, LongCat-2.5 lands between Anthropic's Claude Opus 4.8 and Claude Opus 5 on several mainstream tests.

Read at Sina Techstate media ↗·Read at Leiphone ↗

Apple and Alibaba set December China launch for Qwen powered iPhone AI features

Apple and Alibaba said Alibaba's Qwen model will be built into Apple Intelligence, Apple's suite of AI features, for users in China, with developer testing starting Nov. 4, following the regulatory approval CTD reported October 15th. iPhone users in China will get Apple's AI features with a Chinese model answering some of their requests inside Siri and Apple's writing and camera tools. Users must consent before a question, document or image goes to Qwen, and Alibaba says it will not use those requests to train its models. The features reach Chinese users with the iOS 27.3 update in December, without Apple's Image Playground image tool, per Bloomberg.

Read at PingWest ↗·Read at Tech Wire Asia ↗

AI firm ModelBest's agent writes faster chip code than DeepSeek's handwritten version

Chinese AI company ModelBest's ForgeKernel agent wrote a kernel, the low-level code that runs one math operation on a chip, that beat DeepSeek's handwritten version by 9.8% on Nvidia's B200 chip, QbitAI reports. Kernels largely set how fast an AI model runs, and the team says its method gains most on Chinese chips, whose kernel libraries are less mature than Nvidia's. Huawei's software team for its Ascend AI chips said ForgeKernel topped three of its kernel performance leaderboards, with speedups over baseline of up to 1.9 times. ModelBest has open-sourced the agent, which built the B200 kernel from only the operation's mathematical definition and the chip's hardware constraints in a 22.5-hour livestreamed run.

Read at QbitAI ↗

AI & FOUNDATION MODELS

AI lab DeepSeek publishes design for self-starting coding agent and open-sources a research prototype

Chinese AI lab DeepSeek published a design for an agent that finds bugs and submits code fixes unprompted, and released a research prototype, dsh-superagent, which it says is not an official product. The prototype acts only when asked or when a rule written by the team allows it, and DeepSeek says some designs in its article are not implemented in the code. DeepSeek says one person can typically follow only three to five agent sessions at once, so an agent that starts its own work would let one developer oversee more tasks in parallel. In the design, a coordinating "Router" hands tasks to parallel "Worker" sessions, while a "Watchdog" AI checks whether any task is stuck in a loop or has drifted from its goal. In DeepSeek's simulated tests, 95% of tasks finished within 15.7 minutes, against 29.5 minutes for a conventional single-session setup.

Read at QbitAI ↗·Read at GIGAZINE ↗·Read at GitHub ↗

III

CHIPS & SEMICONDUCTORS

Car chip designer Horizon Robotics reportedly enters BYD driver assist tier that ran on Nvidia chips

Horizon Robotics' Journey 6P chip and HSD driver assistance software have entered the midrange B tier of BYD's God's Eye driver assistance lineup, Chinese media report, with the first such car due in November. The B tier had generally run on one Nvidia chip, so Horizon's entry gives it a domestic supplier, though BYD will decide volumes and how they split with its own in-house system. HSD is an end-to-end system, in which one AI model turns sensor input directly into driving commands. Horizon held nearly 32% of the driver assistance chip market for Chinese-brand cars in the first half, just ahead of Nvidia, 36Kr reports.

Read at EV Industry Technology ↗·Read at 36Kr ↗

DIGITAL INFRASTRUCTURE

State electronics group CETC claims first fully domestic ultra-wideband positioning chip

A team at China Electronics Technology Group (CETC), the state-owned defense electronics conglomerate, says its Tanjie 100 is China's first fully domestic ultra-wideband (UWB) positioning chip. Satellite signals do not penetrate rock and concrete, so the team's underground navigation system transmits BeiDou satellite navigation signals from small base stations that ordinary smartphones can use. UWB chips, which measure distance with very short radio pulses, had long been imported, said project lead Yu Baoguo. The project has passed acceptance by the Ministry of Science and Technology. A stationary person underground can be located to within 0.2 meters, and at a coal mine in Inner Mongolia the system shows miners which tunnel and level they are in.

Read at Xinhuastate media ↗