← China Tech Daily
About Editorial Standards Corrections Privacy Policy
Subscribe →
China Tech · Daily chinatechdaily.org
SEPTEMBER 25, 2026
Your morning brief on Chinese technology. Published by the AI Policy Institute
Bottom Line Up Front

China completes fifth dish of FAST array, part of a planned ring of dozens of midsize antennas that chief engineer Jiang Peng says would raise the giant radio telescope's resolution 200-fold. DeepSeek's unannounced Harness AI agent app surfaces among developers in Mac and Windows builds that 36Kr traced to the company's own download server before any official release. Chinese operator Metastone says software lifts inference nearly sevenfold on a server without high-speed chip-to-chip links, though Nvidia's B300 still delivered 4.9 to 5.6 times the tuned machine's input throughput by Metastone's own comparison.

ITOP STORIES
 
China finishes the fifth dish of a planned array meant to sharpen its giant FAST radio telescope's resolution 200-fold
The main structure of a fifth 40-meter dish in FAST's Phase II test array has been completed a few kilometers from China's 500-meter radio telescope. FAST chief engineer Jiang Peng plans dozens of these dishes around the big telescope, linked into a synthesis-aperture array in which many antennas act as one much larger instrument. The finished array would raise FAST's resolution 200-fold and double its sensitivity, Jiang says. He calls it the world's only array built around a giant telescope ringed by midsize ones. The telescope resumed observations today after its first full overhaul since it began operating a decade ago, Global Times reports. The new dish carries a receiver that needs no cryogenic cooling, which Jiang says took several years to develop.
Read at Global Timesstate media ↗
DeepSeek's unannounced desktop app for its Harness AI agent software surfaces on the company's own download server
Mac and Windows builds of a desktop client for DeepSeek Harness, the Chinese AI lab's software for agents that carry out multistep tasks, are circulating among developers. As of 36Kr's report, DeepSeek had posted no download link. The builds' links match the download domain that the client's open-source code names as its update source. The Mac build had passed Apple notarization, a malware check for apps distributed outside Apple's store, under the name Hangzhou DeepSeek Artificial Intelligence Co. Unlike DeepSeek's free web chatbot, the app requires a funded account and real-name verification, or a developer's own paid key for DeepSeek's models. In its programmatic tool calling mode, the model writes one complete program that batches tool calls and summarizes their results instead of running commands over many network round trips. The build 36Kr installed was labeled version 0.1.7-rc.1, and its built-in updater listed rc.2 as the latest version before any official release.
Read at 36Kr ↗
Chinese compute operator Metastone says software lifts AI inference nearly sevenfold on servers without high-speed chip links
Metastone, a Chinese computing capacity operator, says its Meta-Infer software raised DeepSeek-V4.1-Flash's input throughput nearly sevenfold on a server whose accelerator cards connect only through PCIe. PCIe is the standard slot linking cards to a server, so such machines lack the dedicated chip-to-chip links, such as Nvidia's NVLink, that high-end systems use. The engine targets inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training., the work of running a trained model to answer queries, on chips outside the tested lists of mainstream open-source frameworks. It fills in missing low-level code and rewrites communication between cards without changing the model itself. In the same comparison, Nvidia's high-end B300 still delivered 4.9 to 5.6 times the tuned machine's input throughput on the related DeepSeek-V4-Flash model at matched load. On MiniMax's H3 video model, the PCIe server turned out about 67% of a B300 system's hourly text-to-video clips.
Read at QbitAI ↗
IIPOLICY & REGULATION
 
Senate Democrats urge Trump to seek China deal to slow or pause frontier AI
Seventeen Democratic senators, led by Sen. Chris Van Hollen (D-Md.), sent Trump a letter on Wednesday ahead of his summit with Xi, asking him to pursue an AI accord with Beijing, per Politico. The letter calls for "working with China to develop a formal agreement that would slow down, and potentially enact a mutual pause in, frontier AI development." Signers include Sens. Maria Cantwell (D-Wash.), Elizabeth Warren (D-Mass.) and Tim Kaine (D-Va.). "What our two countries develop as guardrails will serve as a standard for how the technology is used worldwide," the senators wrote.
Read at Politico ↗
IIIROBOTICS & AUTONOMOUS SYSTEMS
 
Driverless delivery firm Neolix teams with courier iMile to take its vehicles to Europe, Latin America and Australia
Neolix, a Chinese maker of driverless urban delivery vehicles, has announced a strategic partnership with courier company iMile to deploy unmanned delivery in Europe, Latin America and Australia. The vehicles run at L4 autonomy, meaning they drive themselves with no human operator inside defined areas. Neolix says it has deployed more than 27,000 of them in hundreds of cities across nearly 20 countries. iMile brings its own delivery network in Europe, Latin America and elsewhere. Running around the clock, the vehicles could cut per-order delivery costs by 30% to 50% once scaled, the partners say. They plan to pair the vehicles with last-mile depots for e-commerce parcels, pilot overnight restocking for supermarkets and test temperature-controlled units for fresh and frozen food.
Read at Gasgoo ↗
IVMILITARY WATCH
 
Chinese navy's 48th escort task force begins a five-day port call in Myanmar
Three ships of the Chinese navy's 48th escort task force, the Tangshan, Daqing and Taihu, arrived at Myanmar's Yangon port for a five-day friendly visit, per Xinhua. Escort task forces are the navy's rotating deployments that guard merchant shipping against piracy in the Gulf of Aden. The ships entered port with weapons raised to ceremonial elevation and crews formed up on deck, greeted by more than 300 people including Myanmar military and government officials and Chinese embassy staff. During the visit, the task force commander is scheduled to call on Myanmar's military and government leaders. The two sides also plan professional exchanges, ship tours and a deck reception.
Read at Global Timesstate media ↗
VSTRATEGIC CONTEXT
 
Xi proposes continued AI dialogue with the United States at White House talks with Trump
Chinese President Xi Jinping proposed that China and the United States continue their dialogue on AI to include exchanging views on AI's risks and benefits, jointly preventing its misuse and keeping the technology under human control, per the Xinhua readout of his talks with President Trump. Xi described both countries as AI powers that compete but can cooperate more, and said the right approach is to "draw on each other's strengths, not guard against each other." The talks also covered a new joint arrangement reached by the two sides' trade teams. On Taiwan, Xi said he hoped the U.S. would take what the readout calls the correct position of opposing Taiwan independence and handle the issue with caution. Xi recalled that the two leaders agreed in May 2026 to build a "constructive China-U.S. relationship of strategic stability."
Read at Science and Technology Daily ↗·Read at Chinese Foreign Ministrystate media ↗

Thanks for reading. See you on Monday.— Daniel

Comments or corrections? hello@theaipi.org

 
Editor-in-Chief:Daniel Colson
Managing Editor:Min Goodman-Cheng
Lead Software Engineer:Christopher Käck
Editorial Advisor:Carrie Adams
Research Manager:Philip Wieczorek
Operations Specialist:Joanne Chua
China Policy Fellow:Tyler Liggett
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.