Baidu open-sources a 280-billion-parameter base model | Huawei's first HarmonyOS desktop claims all-domestic parts | Shanghai lab run by AI performs DNA experiments on its own

Your morning brief on Chinese technology.

Bottom Line Up Front

Baidu open-sourced a 280-billion-parameter base model on Hugging Face, and the company says ERNIE-5.5-Flash-Base outscored same-size base models from DeepSeek, Tencent and Xiaomi on knowledge, reasoning, math and coding tests. Huawei launched its first HarmonyOS desktop PC for government and enterprise offices, and the company says the Qingyun HM650 series received the industry's first certification for a 100% domestic sourcing rate. AutoDNA, a lab run by AI, ran DNA experiments on its own across more than 20 instrument types, according to a Nature paper by its Shanghai Jiao Tong University developers. Chipmaker Hygon's accelerator ran Daxiao Robotics' world model at 94.2% of the speed of a mainstream foreign GPU on a text-to-video task, according to the two companies.

Baidu posted ERNIE-5.5-Flash-Base, a 280-billion-parameter large language model, on the Hugging Face model sharing site before the later training that tunes a model for conversation and specific tasks. Outside developers and researchers can now apply their own fine-tuning and reinforcement learning to a Chinese base model of this size. The model splits its network into specialist subnetworks and uses only 14 billion parameters for each word fragment it processes. Baidu says it outscored same-size base models from DeepSeek, Tencent and Xiaomi on knowledge, reasoning, math and coding tests run under one evaluation framework. A post-trained ERNIE 5.5 Flash model will follow, the company says.

Read at ITHome ↗·Read at PingWest ↗

Huawei launches first HarmonyOS desktop PC, claiming fully domestic components

Huawei launched the Qingyun HM650 series, the first desktop to run its in-house HarmonyOS operating system, with what the company says is the industry's first certification for a 100% domestic sourcing rate. Government and enterprise offices, the machine's target buyers, get a desktop whose processor and operating system both come from Huawei. The company says the platform has passed 1 million PCs shipped. Its Kirin X90 processor includes a dedicated unit for running AI models on the machine itself. More than 21,000 applications, including WPS office software and the DingTalk workplace app, have been adapted for HarmonyOS PCs.

Read at ITHome ↗·Read at Sina Techstate media ↗

Lab run by AI at Shanghai university completes multi-instrument DNA experiments on its own

AutoDNA, a lab run by AI and built by a Shanghai Jiao Tong University team, planned and ran DNA synthesis and sequencing across more than 20 instrument types, according to a Nature paper. Biology experiments normally need trained people to set up instruments and run each step, so a system doing both takes over that skilled hands-on work. In one test it stored data in DNA and read it back exactly over about 200 hours, with people mainly clearing waste and restocking reagents and supplies. Stanford's James Zou wrote in an accompanying Nature commentary that AutoDNA fixed eight of 14 hardware errors on its own, while the other six needed human help.

Read at 36Kr ↗

Chipmaker Hygon's AI accelerator reaches 94.2% of a foreign GPU on robot model

One Hygon Shensuan-3 accelerator card ran Daxiao Robotics' Kairos 3.1 world model at 94.2% of the speed of a mainstream foreign GPU on a text-to-video task, the two companies say. Moving an AI model off a foreign GPU usually costs performance. The rise from 62% at first adaptation shows, by the vendors' own measure, how far software tuning narrowed that gap on one task. A world model predicts how a physical scene will change, so a robot can anticipate what an action will do. The companies tuned the model's software on DTK, Hygon's software toolkit for its accelerators, and passed two rounds of compatibility certification. Other workloads still need verification, the report says.

Read at EV Industry Technology ↗

QUANTUM & FRONTIER SCIENCE

China shows model of ship-borne net that catches rocket stages at sea

A model of the ship-borne net built to catch Long March 10B rocket stages at sea went on display at the China International Import Expo in Shanghai, which opened today. Reusing a rocket's first stage requires recovering it intact, and the net is designed to catch a Long March 10B stage in the 70-metric-ton class. The rocket first flew in July and recovered its first stage at sea on that flight. The structure weighs about 5,000 metric tons and can make the catch in moderate seas, according to the exhibit. It sits on a recovery ship that holds its own position.

Read at The Paperstate media ↗

Chinese university team releases open-source AI model for predicting RNA structure

A University of Science and Technology of China (USTC) team released structRFM, an open-source AI model trained on RNA sequences, with its dataset, code and a paper in Nature. An RNA molecule's function depends on its folded shape, so a model inferring structure and function from sequence alone gives other labs a ready tool for that analysis. Training used about 21 million computer annotated pairs of RNA sequences and their secondary structures, the pattern of base pairs formed where a strand folds back on itself. The model comes from the lab of USTC professor Zhou Shaohua and researcher Xiong Peng. They say it reaches international frontier level on multiple tasks.

Read at Guanchastate media ↗

III

DIGITAL INFRASTRUCTURE

Identity registry for AI agents launches in China, with Ant Group helping build first platforms

A platform for registering and verifying the identities of AI agents, software programs that carry out tasks for users across apps and services, was unveiled at the World Internet Conference's Wuzhen Summit. A shared registry gives platforms a common way to check what an agent is and who authorized it before dealing with it. Ant Group, which runs the Alipay payment app, helped build the first operating platforms on the registry. With more than 20 partners, it is also building a protocol to verify an agent's identity, intent and authorization across platforms. Shanghai's Zhilian Pujiang action plan is the pilot program, with several operating platforms and service providers already connected.

Read at Leiphone ↗