China Desk

China News, Summarized 19 Aug 2026 34 stories archived day

Today's brief

Unitree took $48 billion out of a falling market, and its earliest backer says most of this sector dies

Unitree up 460%, Baidu down 12%

Unitree opened at 1,100 yuan. The Hangzhou maker of the backflipping quadrupeds priced its Shanghai STAR Market listing at 150.80 yuan (~$21) and opened 629% above that, briefly worth over 400bn yuan. It closed at 845 yuan, up 460%, near 342bn yuan (~$48bn) — on 2025 revenue of 1.7bn yuan (~$239m).

The market it listed into fell. The Star Market Composite dropped 7.2% and the Shanghai Composite 2.4%; UBTech, the listed Hong Kong-quoted humanoid maker, fell more than 10%. Meituan's 8.7% post-listing stake was worth nearly 30bn yuan (~$4.2bn) at the close, about seventy times its original investment. One stock, not a sector.

Its first institutional backer is not celebrating. Huang Jinping, founding partner of Rongyi Capital and a fifteen-year ZTE veteran who led Unitree's 2021 A round, told 36Kr he walked twenty-odd friendly funds to Unitree back then and not one followed. Run twenty robot firms at a 500bn-yuan market, he says, and maybe three do well, seven get by, the rest die.

His sharper point is about money arriving too easily. Founders whose rounds close without friction overestimate themselves, and then carry a valuation they have to perform against — so they optimise for the quarter instead of the decade. Every stage of company-building, he argues, has to be lived through; capital cannot buy the skipped years.

Nobody is bidding for search. Baidu's ADR fell more than 12% in New York on Tuesday. A commentary account on Jiemian argues the fix is a break-up: search and models in one basket means the AI side can never be financed on its own terms. Second-quarter capex reached 11.4bn yuan (~$1.6bn) against Tencent's 58bn (~$8.2bn); return on invested capital fell from 7% to about 1%.

Go deeper on this section: ClaudeChatGPT

LandSpace landed a steel booster on legs

Zhuque-3 came home at 7:41am. LandSpace launched the reusable Zhuque-3 Y2 at 7:35, put Honghu-3, an internet satellite, into orbit, and soft-landed the first stage at its pad in Minqin County, Gansu — 66 metres, nine methane-oxygen Tianque-12A engines, roughly 390km (242 miles) downrange. China's first land recovery of an orbital-class first stage, and its first on landing legs.

Two ledgers, same rocket. Solidot's Chinese write-up leads with rank: fourth flight-proven partly-reusable orbital rocket after Falcon 9, New Glenn and Long March 10B, and China now the only country to have recovered boosters two different ways — sea net and legs. It calls this simply the second Zhuque-3 flight.

What it leaves out is the first one. On the December 2025 debut the second stage made orbit and the first stage erupted in a fireball on final descent. Western coverage leads with that turnaround, and with steel: the first stainless-steel booster recovered after an orbital launch, while SpaceX's steel Super Heavy still awaits FAA clearance for orbital recovery. SpaceNews adds the fixes — fewer engines in the landing burn, a predicted landing point in the onboard safety logic, better thermal protection.

Cheap mass up does not mean cheap bits up. GeekPark's interview with Space Byte, a satellite data-centre startup co-founded by a Beidou veteran, has the number that matters: uploading an 800MB model to a satellite already in orbit takes close to four weeks. Command links run around 1Mbps, and a pass gives you eight to ten minutes.

Nobody built the uplink because nobody needed it. Satellites were designed to fire data down and receive instructions. Laser downlink records keep falling; laser uplink at low orbit fires through the turbulent air first, and remains a test-bench result everywhere. Thirty-odd Chinese firms now say they do orbital compute. The bottleneck is a fire hose pointing the wrong way.

Go deeper on this section: ClaudeChatGPT

Nobody knows who runs the AI talks

Xi's Washington trip is weeks away and the AI dialogue has no shape. SCMP reports venue, participants and scope still unsettled, with the US side yet to decide which part of government leads and the two sides not agreed on whether technical experts and companies attend at all.

That asymmetry is the story. Beijing arrives with one interlocutor and one position. Washington arrives with an unresolved turf question, which means the first substantive negotiation is internal. Whoever wins it sets the agenda by default.

The provident fund got rewritten for consumption. Li Qiang signed amendments to the housing fund rules effective 20 September, per Sinocism: the rent-to-income threshold for withdrawals is gone, renovation and property-management fees become eligible, and gig and part-time workers may contribute voluntarily. Household stimulus that costs the budget nothing — households spend their own savings.

Jiang Zemin's centenary is being edited as it is celebrated. Cai Qi, the Standing Committee member who runs the Party's Central Committee General Office, spoke on the Three Represents and, in Sinocism's reading, used it to underwrite the 15th Five-Year Plan while leaving out market economy, ownership pluralism, private entrepreneurs and external integration — which is most of what the phrase originally licensed.

Go deeper on this section: ClaudeChatGPT

DeepSeek's harness now ships whole agents

Five days, 149,000 stars. GeekPark's tally for DeepSeek Harness: 15,000 forks, 5,100 plugins, 3,500 authors. The community shipped desktop wrappers within 48 hours, one down to 18MB. DeepSeek then published a dozen of its own internal engineering Skills — code review, dead-code hunting, stripping leftover chain-of-thought from docs.

The new unit is the Preset. Not a skill, not a plugin: a bundle of plugins, skills, browsing, context policy and loop, packaged and distributed as a finished agent. Sites now exist to hand out research agents and video agents in one click. The harness stopped being the product; the agent is.

Meanwhile the layer commoditises. Plugins route to GLM, Kimi, Qwen, Claude and Gemini; one vision plugin covers fifteen-plus providers because V4-Pro still cannot see. DSH runs as a subagent inside Codex and Claude Code, and vice versa. WeChat Work opened its CLI and MCP to any agent, no company-size threshold.

And the man who argued this leaves MiniMax. Miao Yuhang, its head of engineering across M3.x, Code, Audio and the Hailuo consumer app, is marked as departed with no next stop announced. His public line was that a model is an F1 car and the harness is the driver — and that whatever the harness adds today, the model eventually swallows.

Go deeper on this section: ClaudeChatGPT

A Qianmen travelogue got deleted anyway

The post described a security queue and nothing else. China Digital Times translated a WeChat account's account of visiting Zhengyangmen at the south end of Tiananmen Square: ID swipes, bag searches, book checks, hems lifted, shoes squeezed, in stifling heat. No criticism, no history, no complaint — just advice to avoid August.

It was removed for violating the Cybersecurity Law. Which is the tell. The threshold has moved from opinion to description: writing down what the queue is like is now sufficient. Keyword scanning may have done it automatically, which is worse — nobody decided.

Hong Kong is running out of adjectives. Initium reports 36.9°C on 9 August, the highest since 1884, and a night that stayed above 30.2°C, the warmest in 143 years. Three suspected heat deaths in a week: two men on hiking trails, a decorator found dead in a Tai Po flat after telling his wife he was exhausted.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The grey market underpriced Unitree by half. Brokers bid 520 yuan (~$73) pre-listing and crypto perpetuals implied a 300% pop; it closed at 845 and touched 1,100. Founder Wang Xingxing is worth over $12bn on paper. The debut did not lift the rest of the sector, and DeepRobotics and LimX are next in the queue.
  • Wang used the listing lunch to name his bet. GeekPark's account: a loop where AI reads papers, writes training code, runs simulation, deploys to real hardware, scores the result and retrains. The moat is the last step — 33,000 quadrupeds and 5,500 humanoids shipped. 2.02bn yuan (~$285m) of proceeds goes to robot models, the largest single line.
  • Geely's handover is done. An Conghui took the chair on 18 August; H1 revenue 173.6bn yuan (~$24.5bn), exports up 158%, and no further combustion-engine development. The outgoing chief executive called the results "eye-catching but not amazing".
  • New watch — Nokia's China exit. It plans to cut most of its China staff by year-end and close plants in phases, keeping after-sales; severance is N+3. It reported about 7,200 Greater China employees at end-2025, with a Beijing R&D centre dating to 1995.
  • A broker put Qwen's harness first. Jefferies tested eight agents on five office tasks; Alibaba's Qwen Office scored above 90 on every dimension, ahead of Claude Cowork and Codex, with the highest implied harness score. Reported through a Chinese outlet, framed generously — but cost per task is the right metric.
  • Entanglement reached 420km. USTC linked two cold-atom quantum memories over 420km of fibre and beat the no-repeater limit above 230km, up from 50km in 2020.

Go deeper on this section: ClaudeChatGPT

The river · 34

Unitree's founder outlines a self-improving robot development loop

# GeekPark

At Unitree's IPO banquet, founder Wang Xingxing described a system that connects model search, code generation, simulation, real-robot tests and evaluation into a physical-AI flywheel.

More
  • The proposed loop would let AI search papers and code, generate training programs, run simulations, deploy them to robots, score real-world behavior and feed the results into the next iteration.
  • Unitree's distinctive asset is not only a model but a deployed hardware base: the company reports more than 30,000 cumulative quadruped sales and over 5,500 humanoids shipped in 2025.
  • That installed base could supply real-world feedback at a scale unavailable to companies still limited to prototypes, provided users authorize data collection and the data is standardized.
  • The founder's thesis is that model and coding capabilities will diffuse quickly, so the durable moat will be the ability to run more physical experiments cheaply and safely.
  • Early versions would still need humans to define tasks, success criteria and safety boundaries; the claim that the system could outperform a hundreds-person AI team remains a future scenario.
  • The architecture resembles automated research harnesses, but physical testing introduces failures, maintenance, hardware variance and liability that cannot be simulated away.
  • If Unitree turns its manufacturing scale into a trusted data network, the IPO could finance a development platform rather than merely a robot catalog.
Open at GeekPark Ask Claude ChatGPT

A deleted post describes extreme security around Beijing’s Qianmen

# China Digital Times

A deleted Chinese post describing identity scans, searches, and intrusive checks at Beijing’s Qianmen was removed under cybersecurity rules, showing how censorship can reach mundane travel advice.

More
  • The original account reportedly described ordinary visitor procedures without overt political criticism, yet the post was deleted and replaced by a notice citing China’s Cybersecurity Law.
  • The significance is collateral censorship: automated or complaint-driven enforcement can suppress neutral descriptions of state security practices even when they contain no sensitive historical discussion.
  • The checks included identity verification, bag and book searches, pat-downs, and inspection of clothing and shoes, according to the archived translation.
  • This is a useful example of how security around politically symbolic sites becomes normalized in daily life while public discussion of the experience remains risky.
  • The source is an archived translation of deleted material, so the exact enforcement trigger—human complaint, keyword matching, or both—is unclear.
  • The broader governance pattern is that content moderation protects not only political secrets but also the unremarkable visibility of state security.
Open at China Digital Times Ask Claude ChatGPT

LandSpace recovers the first stage of its Zhuque-3 rocket

# SCMP

LandSpace successfully placed a satellite in orbit and soft-landed Zhuque-3's first stage, giving China a flight-validated, land-recoverable orbital booster after the Falcon 9 model.

More
  • The 66-meter rocket launched from Jiuquan and landed about eight minutes later in Gansu after delivering the Honghu-03 satellite.
  • The report identifies nine Tianque-12 engines on the first stage and one Tianque-15 engine on the second stage, with a landing approach broadly similar to Falcon 9.
  • It is described as China's first land recovery of an orbital-class first stage using landing legs; an earlier Long March 10B demonstration used a sea-based net recovery.
  • Recovery is only the beginning of reusability: the hard economic test is inspection, refurbishment, turnaround time and repeat flight with a useful payload.
  • A successful booster landing lowers launch costs only when cadence rises and hardware survives multiple flights without expensive bespoke work.
  • The result sharpens competition between China's state launch programs and private companies such as LandSpace over reusable propulsion, launch infrastructure and satellite customers.
  • The next decisive milestone is a reflown stage and a sustained launch schedule, not the spectacle of a single landing.
Open at SCMP Ask Claude ChatGPT

China's Zhuque-3 first stage completes a land landing

# Solidot

LandSpace successfully launched Zhuque-3, placed a satellite in orbit and landed its first stage on land, making it China's first orbital-class stage recovered by landing legs.

More
  • The rocket lifted off at 7:35 and landed its first stage six minutes later at the Zhuque-3 landing site in Gansu, according to the report.
  • The vehicle is 66 meters tall and 4.5 meters wide, using nine Tianque-12 engines on the first stage and one Tianque-15 on the second.
  • The mission was Zhuque-3's second flight and reportedly delivered Hongqing Technology's Honghu-03 satellite to its planned orbit.
  • The landing makes Zhuque-3 the fourth flight-proven partially reusable orbital rocket type after Falcon 9, New Glenn and China's Long March 10B.
  • The engineering significance is repeatability: a second flight and a controlled land recovery move China closer to a high-cadence launch model rather than a one-off demonstration.
  • The next proof points are stage refurbishment, reflight, turnaround time and whether the recovery penalty can be justified by launch economics.
  • This is a strategic infrastructure milestone because reusable launch lowers the marginal cost of deploying large satellite constellations and increases pressure on China's launch ecosystem to industrialize.
Open at Solidot Ask Claude ChatGPT

Xiaomi crosses 100 billion yuan in quarterly revenue as it funds AI and cars

# Leiphone

Xiaomi reported 108.9 billion yuan (about $15.3 billion) in second-quarter revenue, a record 1,351 yuan phone ASP and 104,199 EV deliveries while raising R&D spending 18.9%.

More
  • Operating profit reached 10.87 billion yuan (about $1.5 billion), up 104.6% sequentially, while phone ASP rose 25.9% as the company pushed into higher-priced models.
  • EV and innovation revenue reached 24.9 billion yuan (about $3.5 billion), with deliveries up 28.2%; Xiaomi says cumulative SU7 deliveries passed 500,000 by August 17.
  • Second-quarter R&D spending was 9.2 billion yuan (about $1.3 billion), with AI, operating systems, chips, robotics and vehicles absorbing investment while the core phone business funds it.
  • The business is becoming a portfolio rather than a handset maker: hardware margins, an installed IoT base and software services can subsidize longer-cycle bets.
  • Xiaomi says its AIoT platform connects 1.16 billion devices and that 24.6 million users have five or more connected devices, creating distribution for future agents and services.
  • The key risk is execution across too many capital-intensive categories at once, especially with memory costs high and Chinese consumer demand uneven.
  • If Xiaomi can turn its devices, cars and operating system into one software surface, its ecosystem may be more defensible than any individual product; if not, the investment load will dilute the phone business.
Open at Leiphone Ask Claude ChatGPT

Beijing blocks Chinese firms from helping the EU probe JD.com

# SCMP

China's Ministry of Justice ordered Chinese entities not to assist an EU foreign-subsidy investigation into JD.com's proposed acquisition, turning a competition case into a sovereignty dispute.

More
  • Beijing called the EU investigation an improper extraterritorial measure and said no organization or individual may execute or assist it.
  • The EU probe concerns JD.com's proposed takeover of German electronics retailer Ceconomy, according to the report.
  • The order creates a direct compliance conflict for companies, auditors and advisers that may face obligations under both Chinese law and EU investigative rules.
  • The measure signals that Beijing is willing to use blocking authority to defend Chinese companies against foreign subsidy scrutiny, not only to manage data or export controls.
  • For future Chinese acquisitions in Europe, due diligence will have to model regulatory discovery and evidence-sharing risk as part of the transaction itself.
  • The immediate practical question is whether the EU escalates against JD.com or intermediaries, and whether Beijing's order becomes a template for other foreign investigations.
  • This is a structural clash over whose regulator can compel information from a cross-border deal, not merely a disagreement about JD.com's market conduct.
Open at SCMP Ask Claude ChatGPT

China and the US still lack a concrete agenda for planned AI talks

# SCMP

With Xi Jinping's planned US visit approaching, the venue, lead agency, participants and scope of bilateral AI talks remain unsettled, showing how little institutional machinery exists for the dialogue.

More
  • The supplied report says Washington has not decided which part of the government should lead and that the two sides have not agreed on technical or private-sector participation.
  • That uncertainty matters because AI diplomacy spans export controls, safety evaluation, compute, model access and commercial competition, each owned by different US institutions.
  • A leader-level announcement without an agreed working channel risks producing symbolism rather than durable risk-management mechanisms.
  • China's likely preference is a state-to-state framework that treats AI as a strategic technology and sovereign capability; US agencies and companies may seek a more technical or operational format.
  • The absence of basic arrangements weeks before a planned visit suggests that strategic mistrust is limiting even low-level agenda-setting.
  • A useful dialogue would need narrow deliverables, such as incident communication, evaluation terminology or research-contact rules, rather than an attempt to resolve the entire technology relationship.
  • Whether the talks matter will be visible in who attends and whether the meeting creates a continuing channel below the leaders.
Open at SCMP Ask Claude ChatGPT

China plans a marine monitoring station near Scarborough Shoal

# SCMP

Chinese authorities approved a marine observation station covering waters around disputed Scarborough Shoal, presenting a scientific facility that also strengthens Beijing's administrative presence.

More
  • The station was proposed by the Chinese Academy of Sciences' South China Sea Institute of Oceanology and approved on July 31, according to the supplied report.
  • Its stated mission is environmental observation and research, but the location gives China a durable platform for collecting data and maintaining institutional presence in contested waters.
  • Marine monitoring infrastructure can support fisheries, environmental enforcement, navigation and legal claims, so its strategic value exceeds the scientific label.
  • The move fits Beijing's pattern of using civilian and bureaucratic instruments to normalize control without announcing a new military base.
  • For the Philippines, the issue is not only the facility itself but whether Chinese research and monitoring activity changes the practical operating environment around the shoal.
  • Chinese coverage is likely to stress ecological stewardship and administrative legitimacy; regional critics will read the same infrastructure as incremental consolidation.
  • The effect will depend on construction, staffing, data publication and whether Chinese agencies use the station to support patrols or regulatory actions.
Open at SCMP Ask Claude ChatGPT

Chinese team entangles quantum memories across 420 kilometers

# Solidot

A University of Science and Technology of China team reports entanglement between cold-atom quantum memories over 420 km of fiber, extending the building blocks of an intercity quantum network.

More
  • The experiment exceeded 230 km, the report's stated theoretical limit for direct entanglement distribution without a repeater.
  • The memories are physical quantum nodes rather than only photons, making the result relevant to repeater-based communication and distributed quantum computing.
  • The team reports a progression from 50 km in 2020 to a three-node metropolitan network in 2024 and high-fidelity hundred-kilometer links in 2026.
  • The distance is a network milestone, not a complete quantum internet: fidelity, rate, memory lifetime and repeater operations determine practical usefulness.
  • The result strengthens China's position in long-distance quantum networking, an area where sustained system engineering matters as much as a single laboratory distance record.
Open at Solidot Ask Claude ChatGPT

Geely founder Li Shufu gives way to veteran executive An Conghui

# Jiemian

Geely Auto has replaced founder Li Shufu as chairman with longtime executive An Conghui, signaling a shift toward professional management as the company pivots fully to hybrids and overseas growth.

More
  • Li remains chairman of the larger Zhejiang Geely Holding Group and becomes Geely Auto's lifetime honorary chairman, separating group ownership from listed-company execution.
  • The reshuffle also moves Gui Shengyue to vice chairman and promotes Gan Jiayue to chief executive, suggesting a broader succession plan rather than a single title change.
  • Geely Auto reported first-half revenue of 173.6 billion yuan (roughly $24.5 billion), up 15%, while attributable profit fell 2% to 9.091 billion yuan (roughly $1.28 billion).
  • Management said the company will stop developing conventional gasoline vehicles and move its remaining combustion models toward intelligent hybrid technology.
  • The overseas sales target was raised to 920,000 vehicles, making execution outside China a central test for the new leadership.
  • The strategic question is whether a more professional structure can preserve Geely's multi-brand engineering discipline while reducing founder dependence.
Open at Jiemian Ask Claude ChatGPT

China rewrites housing-fund rules to unlock idle savings

# Jiemian

A revised housing-fund regulation broadens withdrawals beyond home buying and targets roughly 11 trillion yuan (about $1.55 trillion) in accumulated balances.

More
  • The new framework covers renting, repairs, renovation and some retirement-related needs, shifting the fund from a purchase subsidy toward a broader housing-support system.
  • The regulation is scheduled to take effect on September 20, 2026, according to the report.
  • The change reflects China’s move from property-led expansion toward managing existing housing stock and household affordability.
  • It also responds to a low-rate environment in which mandatory balances are less effective as a housing-purchase lever and remain politically visible as idle funds.
  • Broader withdrawals could support consumption and home maintenance, but they may also reduce the pool available for mortgage-linked financing.
  • The crucial implementation question is whether local administrators can expand access without weakening fund liquidity or creating new regional disparities.
Open at Jiemian Ask Claude ChatGPT

Unitree surges 629 percent in its Shanghai debut

# IT Home

Unitree opened on the Shanghai exchange 629 percent above its issue price, putting China’s humanoid-robotics enthusiasm on public display even as earnings lag growth.

Also: Techmeme

More
  • The stock opened at 1,100 yuan and briefly implied a market value of 444.9 billion yuan, roughly $62.7 billion, against an issue price of 150.80 yuan.
  • Unitree reported first-half revenue of 1.152 billion yuan, roughly $162 million, up 48.5 percent, while adjusted net profit fell 19.3 percent to 244 million yuan, about $34 million.
  • The company sells quadruped and humanoid robots and claims global leadership in shipments of both categories, though the item provides no independent market validation.
  • The extreme opening move reflects scarce public robotics exposure and retail speculation as much as operating performance.
  • For engineers and executives, the important question is whether Unitree can convert hardware shipment leadership into repeatable software, service and industrial revenue.
  • A small freely tradable float can amplify the first-day move, so the debut is a weak proxy for the company’s fundamental value.
Open at IT Home Ask Claude ChatGPT

LandSpace recovers a stainless-steel booster after reaching orbit

# IT Home

LandSpace says Zhuque-3 reached orbit and softly landed its first stage, while Chinese coverage stresses that reusability—not launch alone—is the next test against SpaceX.

More
  • The rocket uses a stainless-steel body and liquid-oxygen methane engines, a design path closer to Starship than Falcon 9 and potentially cheaper at scale.
  • Deputy chief designer Kai Dong says the immediate goal is to reuse the recovered stage, then shorten turnaround below the time needed to build a new vehicle.
  • The booster is currently at flight-validation level seven on a nine-level engineering scale, far behind Falcon 9 in operational maturity.
  • The first stage uses nine variable-thrust Tianque-12A engines and is designed for vertical recovery, while a future engine upgrade is intended to raise payload capacity.
  • LandSpace wants to increase Zhuque-3 operations to 30 to 50 launches, but larger rockets remain contingent on demand and the performance of this fleet.
  • The meaningful milestone is not the landing itself but whether inspection, refurbishment and repeat launches can become an industrial process.
Open at IT Home Ask Claude ChatGPT

Z.ai launches GLM-5.3 API and plans to open its weights

# IT Home

Chinese coverage sells GLM-5.3 as a low-cost frontier model, while US tech discussion focuses on its benchmark tie with Kimi K3 and coming open-weight release.

More
  • Z.ai says GLM-5.3 scores 60 on the Artificial Analysis Intelligence Index and keeps the previous model’s API pricing, positioning post-training gains as the main improvement.
  • The model uses the same base model as GLM-5.2, with capability gains attributed mainly to reinforcement-learning-style post-training rather than a new pretraining run.
  • The company says the model targets complex coding, defensive cybersecurity and long-horizon tasks, and plans to release weights the following Friday.
  • The Chinese framing emphasizes intelligence per unit cost and broader access, while the US aggregator framing treats the score and open-weight ranking as the news.
  • Because the claims are company-reported and benchmark-dependent, the important test is whether independent users reproduce the cost-performance advantage after release.
Open at IT Home Ask Claude ChatGPT

Shanghai shuts a 30-storey ghost-delivery operation

# IT Home

Chinese coverage stresses food-safety enforcement in a Shanghai case where one barbecue shop used forged licenses to run more than 30 delivery storefronts across seven districts.

More
  • Investigators say the shop copied or digitally altered real business licenses, then used them to pass platform checks and collect orders under addresses that did not prepare the food.
  • The actual kitchen was more than 10 miles from the listed Huangpu storefront, creating a direct mismatch between platform identity, delivery geography and food accountability.
  • The operation generated tens of thousands of yuan per month and more than 1 million yuan, roughly $141,000, overall, according to the report.
  • The case shows why platform-level merchant verification is a public-health control, not merely an onboarding inconvenience.
  • Authorities have transferred the suspects to police on alleged forgery of official documents; the case remains under investigation.
  • The report places the case beside a national crackdown in which seven platforms were fined a combined 3.597 billion yuan, roughly $507 million, showing that ghost kitchens are being treated as a systemic platform problem.
Open at IT Home Ask Claude ChatGPT

China prepares Chang'e 7 mission to search for lunar water ice

# IT Home

Chang'e 7 has moved to the launch area for a lunar south-pole mission built around landing, hopping and probing permanently shadowed craters for water ice.

More
  • The mission targets the region south of 85 degrees latitude in the Aitken Basin, where permanently shadowed areas may preserve ancient volatiles.
  • Its planned capabilities include precision landing, legged surface travel, hopping and in-situ work inside permanently shadowed craters.
  • The scientific payloads will study surface conditions, regolith water and volatiles, geology, shallow structure, moonquakes, magnetic fields and the space environment.
  • The mission also carries international payloads, showing that China's lunar program can combine national control with selective scientific cooperation.
  • The strategic value is not just finding water; a successful south-pole operation would validate navigation, thermal management, communications and autonomous mobility needed for later infrastructure.
  • The supplied report frames Chang'e 7 as preparation for Chang'e 8 and a basic international lunar research station, making it part of a staged program rather than a standalone science flight.
  • The decisive evidence will be actual resource measurements and repeatable surface operations, not the launch narrative.
Open at IT Home Ask Claude ChatGPT

Yangtze Memory moves closer to an A-share IPO

# IT Home

China’s securities regulator has moved Yangtze Memory’s IPO coaching status to final acceptance, a step toward filing that matters for the country’s NAND independence and capital access.

More
  • The status ends the preparatory coaching phase but does not mean the company has passed the listing review or will be approved to trade.
  • Yangtze Memory is described as mainland China’s only dedicated 3D NAND integrated device manufacturer, serving consumer, server and data-center markets.
  • The company was valued at 1.6 trillion yuan, roughly $225 billion, in a 2025 unicorn ranking, making the eventual filing a major test of semiconductor-market appetite.
  • Counterpoint’s second-quarter 2026 tracker reportedly placed Yangtze Memory at 14 percent of NAND share, its first appearance in third place.
  • An IPO would provide a domestic financing channel for capacity, process development and supply-chain resilience while exposing the company to much greater disclosure.
  • The strategic significance is larger than the procedural milestone: a public listing would formalize a key pillar of China’s memory strategy, subject to the regulator’s later scrutiny.
Open at IT Home Ask Claude ChatGPT

China launches a high-throughput satellite technology demonstrator

# IT Home

The Honghu 03 satellite reached orbit on Zhuque-3 with an argon electric thruster, active thermal control and a claimed terabit-per-second onboard router.

More
  • The satellite is described as a technology demonstrator for future satellite-internet constellations, not a commercial service unit.
  • Its technical payload includes an argon Hall thruster, an active fluid-loop radiator and a high-performance onboard electronics system.
  • The reported router is intended for tests spanning inter-satellite transmission, satellite-ground networking and combined communications and computing.
  • The launch is presented as the first satellite built in Xiong'an to move from local research and manufacturing into orbit.
  • If these subsystems mature, the significance is manufacturing repeatability: China wants to move from single demonstrators to batch production for a domestic constellation.
Open at IT Home Ask Claude ChatGPT

Chinese team uses a dark cavity to raise a superconductor's critical temperature

# IT Home

A Chinese-led team reports the first experimental enhancement of superconductivity through vacuum fluctuations, raising a six-layer device's critical temperature by up to 5.4%.

More
  • The experiment placed niobium diselenide inside a terahertz split-ring resonator that reshapes the local electromagnetic environment without external photons.
  • The reported maximum increase in critical temperature was 5.4%, accompanied by higher critical current and critical magnetic field near the transition.
  • A resonance-shaped dependence on cavity frequency, plus control experiments, is presented as evidence that the effect is coupling rather than strain or sample degradation.
  • The proposed mechanism is virtual-photon exchange between the superconducting state and the cavity mode, which lowers the state's energy when the frequencies match.
  • This is a laboratory demonstration of a non-contact control knob, not a practical high-temperature superconductor or a device-ready architecture.
  • The result matters because it turns engineered electromagnetic boundary conditions into a possible materials-design parameter.
  • Its significance will depend on replication, absolute temperatures, effect size in other materials and whether the cavity can be integrated without sacrificing useful current or magnetic-field performance.
Open at IT Home Ask Claude ChatGPT

Chinese lab deploys humanoid robots for autonomous experiments

# AI QbitAI

A Chinese national research laboratory is using Monte2 robots for long sequences of liquid handling, cell culture and toxicity testing, moving embodied AI from demos into scientific operations.

More
  • The robots reportedly perform reagent transfer, preparation, pipetting and instrument coordination without continuous human operation.
  • The system is built around an object-centered physical model that tracks samples, reagents, consumables, instruments and their relationships rather than executing one fixed motion.
  • The reported capability includes hours of continuous operation and sub-millimeter precision while adapting to liquid levels, tube angles and grasping forces.
  • The laboratory plans to expand the fleet to roughly 100 robots by the end of 2027 and coordinate them through a common AI system.
  • This is a valuable deployment domain because experiments generate measurable outcomes and repeated procedures, creating a natural feedback loop for models.
  • The important missing evidence is throughput, contamination rate, reproducibility against trained technicians and the cost of maintaining a humanoid platform.
  • If validated, the lab becomes both a customer and a data engine for physical AI, with implications for drug discovery and national research infrastructure.
Open at QbitAI Ask Claude ChatGPT

Embodied-AI data platform goes on sale for about $720

# AI QbitAI

51WORLD launched a hardware-software platform that captures human demonstrations, filters bad data at the edge and feeds validated trajectories into robot training and simulation.

More
  • The first AperData kit is priced at CNY 5,100, about $720, and combines a multi-camera capture device with a data-governance platform.
  • Its four-camera configuration reports more than 270 degrees of horizontal view, hardware timestamp synchronization and trajectory physical consistency up to 99%.
  • The important workflow change is capturing people directly rather than occupying expensive robot bodies with teleoperation during every demonstration.
  • The platform screens blur, exposure, dropped frames, synchronization and incomplete tasks at the edge, reducing wasted upload, storage and cloud-compute costs.
  • The vendor claims more than a tenfold production-efficiency improvement over traditional robot teleoperation at equivalent cost, a figure that needs independent field validation.
  • A second platform links reconstruction, training, evaluation, deployment and operations, showing the emerging market for embodied-AI infrastructure below the model layer.
Open at QbitAI Ask Claude ChatGPT

A Tencent recommendation contest tests a unified architecture

# AI QbitAI

A Tencent-backed KDD Cup challenge drew 13,913 contestants from 52 countries and found that a unified recommendation block could raise ranking quality while cutting inference cost.

More
  • The industrial winning system reportedly improved AUC by 0.0048 while reducing per-sample compute by 18%, a modest-looking change that scales materially at billion-request volumes.
  • The winning academic design used an intent token to carry user-interest information across layers while allowing profile and item features to steer its updates.
  • The competition reflects a systems problem distinct from language models: recommendation inference must meet millisecond-scale latency and enormous request volume.
  • Its 6-million-yuan-plus prize pool and global participation show Chinese internet companies using competitions to expose real, anonymized production problems to international researchers.
  • The result is not proof of a universal recommendation scaling law, but it offers a concrete architecture and efficiency tradeoff worth testing in production.
Open at QbitAI Ask Claude ChatGPT

Humanoid robot plays a full autonomous table-tennis match

# AI QbitAI

SuperDynamix demonstrated a humanoid system that perceives, predicts, plans and controls a complete table-tennis exchange, using a world model and 117-degree-of-freedom body.

More
  • The SMASH 2.0 system must identify a fast, spinning ball, predict its trajectory and coordinate shoulder, elbow, wrist, torso and legs within millisecond-scale windows.
  • The company says the same control algorithm can run on different robot bodies, a meaningful claim about policy portability rather than just athletic performance.
  • Its KAI world model is pretrained on millions of videos and is intended to generate or reconstruct environments for simulation and closed-loop reinforcement learning.
  • The 37-degree-of-freedom hand and first-person KAI Halo capture system show the company's attempt to control both action hardware and data production.
  • The demonstration is technically richer than a scripted swing, but the commercial significance depends on whether the stack transfers to manipulation, not whether it wins a rally.
Open at QbitAI Ask Claude ChatGPT

RL infrastructure becomes the bottleneck for China's reasoning models

# AI Leiphone

A technical analysis of Zhipu and Jiuzhang argues that reinforcement learning turns model training into a continuously balanced production system linking rollout generation, environments and weight updates.

More
  • The core loop is generate, reward, train, update and generate, with the generator and trainer sharing weights, rollouts, KV cache, environment state and task queues.
  • If the generator is slower than the trainer, GPUs idle; if it is faster, rollouts accumulate and become stale, creating a distributed-systems problem rather than a simple training problem.
  • The analysis emphasizes matching trainer throughput with effective generator throughput and keeping policy updates synchronized with long agent trajectories.
  • It cites experiments in which smaller reasoning models improved through continued on-policy RL, while rewarding longer answers alone led to length hacking.
  • The architectural lesson is that post-training scaling requires scheduling, state management, fast inference and environment execution as much as it requires a better reward model.
  • The article uses GLM examples and vendor infrastructure claims, so performance numbers should not be treated as independent benchmarks.
Open at Leiphone Ask Claude ChatGPT

Chinese coastguard journal proposes robotic South China Sea outposts

# Policy SCMP

A Chinese defense proposal envisions autonomous systems and machine swarms defending disputed islands, revealing how Beijing is thinking about persistent, low-cost maritime control.

More
  • The proposal reportedly replaces or supplements soldiers with coordinated machines designed to counter low-cost drone swarms and other asymmetric threats.
  • A robotic outpost could combine sensors, communications, unmanned surface or aerial vehicles and remote command, extending presence while reducing personnel exposure.
  • The concept fits a broader Chinese military trend toward distributed systems and autonomous cooperation rather than a few expensive platforms.
  • Because the proposal appeared in a journal supervised by a state shipbuilding group, it signals an institutional research direction but not an approved deployment plan.
  • The South China Sea setting matters: autonomous infrastructure could make contested features more continuously observable and harder for rivals to challenge without escalation.
  • The unresolved questions are communications resilience, rules for machine-initiated force, maintenance in harsh maritime conditions and control during spoofing or jamming.
  • Even if never built as described, the proposal shows how China is translating commercial robotics and drone technology into gray-zone maritime strategy.
Open at SCMP Ask Claude ChatGPT

China recovers orbital boosters through two different technologies

# Analysis SCMP

LandSpace's successful booster recovery makes China the first country to validate two distinct orbital-recovery paths, combining a new land landing with an earlier sea-based recovery approach.

More
  • The recovered stainless-steel first stage followed a conventional powered landing rather than a net capture, while the Long March 10B has demonstrated a sea-based net system.
  • Different recovery architectures let China test tradeoffs among landing infrastructure, vehicle design, weather tolerance and launch-site geography.
  • The milestone strengthens the case that Chinese commercial and state launch programs are pursuing reusable operations in parallel rather than relying on one national design.
  • The engineering achievement is significant, but recurring launches and reuse of recovered stages are still needed before the cost advantages become commercial reality.
Open at SCMP Ask Claude ChatGPT

A hit Iraq-war film reveals China's preferred view of intervention

# Culture SCMP

A Chinese film about the Iraq war uses a Chinese chef's perspective to frame US intervention through ordinary civilian experience, extending Beijing's soft-power effort to shape how foreign audiences read its worldview.

More
  • The film avoids naming Iraq directly but depicts an invasion by American troops, allowing the political message to travel through allegory rather than an explicit state argument.
  • Centering a Chinese worker abroad gives the story an emotional bridge between China's own citizens and the consequences of great-power intervention.
  • The production is part of a broader cultural strategy in which historical and foreign conflicts are used to present China as an observer, victim or alternative source of order.
  • Its influence will depend on distribution and audience reception outside China; a state-aligned message does not automatically become effective soft power.
  • The film matters less as a literal account of the Iraq war than as evidence of how Chinese popular culture is being used to narrate American power and global instability.
Open at SCMP Ask Claude ChatGPT

DeepSeek Harness grows from rough toolkit into agent marketplace

# AI GeekPark

DeepSeek Harness reportedly reached 149,000 GitHub stars and 5,100 plugins in five days, as its modular architecture let the community repair its rough edges.

More
  • The project's central design makes nearly every subsystem replaceable, including model adapters, tools, file systems, sandboxes, scheduling, and even the agent loop.
  • That creates a trade-off: the initial product can feel unfinished, but external developers can add a desktop shell, image support, IDE features, and migrated workflows without waiting for the core team.
  • The architecture treats logs as the durable center, making model-visible context reconstructable and potentially improving auditability across interchangeable plugins.
  • The rapid plugin growth is a community metric, not evidence that the extensions are secure, reliable, or suitable for enterprise use.
  • The strategic lesson is that open agent systems can outsource product completion to developers, but they also shift permission, supply-chain, and maintenance risks into the ecosystem.
Open at GeekPark Ask Claude ChatGPT

Baidu says embodied AI is moving from demos to measurable work

# AI Leiphone

Baidu argues that robots will commercialize through a model-data-application loop, beginning with bounded industrial tasks whose success rate, takeover rate and cost can be measured.

More
  • Baidu's cloud platform is intended to cover data processing, model training, simulation, evaluation and real-robot inference, with some reported single-step latency below 100 milliseconds.
  • The company identifies industrial loading, material handling, inspection and patrol as earlier entry points than full job replacement because their outputs are easier to measure.
  • Baidu emphasizes failure, correction, human takeover and boundary cases as more valuable training data than repeated successful demonstrations.
  • The strategic message is that embodied AI infrastructure is becoming a systems business: cloud, data governance, simulation, security and deployment all matter alongside the policy model.
  • The source is a conference statement from a company executive, so the latency and deployment claims need customer-level verification.
Open at Leiphone Ask Claude ChatGPT

Chinese fighter designers warn of hallucinations in military AI

# Analysis SCMP

Engineers associated with China's J-20 and J-36 programs warn that language models can invent aircraft specifications, offering a rare Chinese military critique of AI reliability.

More
  • The engineers say large language models can rapidly mine intelligence and analyze weapons performance, but may fabricate technical specifications.
  • The warning is especially important in defense because a plausible but false aircraft parameter can contaminate threat assessment, simulation or procurement decisions.
  • The paper's authors work at the Chengdu aircraft design institute associated with the J-20 and next-generation J-36 programs, according to the report.
  • The Chinese framing is cautious rather than triumphalist: AI is useful for accelerating analysis, but human verification remains mandatory in high-consequence domains.
  • The report does not describe a deployed military system or a specific incident caused by a hallucination.
Open at SCMP Ask Claude ChatGPT

SynapX presents a brain-hand-data stack for physical AI

# AI QbitAI

SynapX's World model, high-degree-of-freedom hand and electromyography-based data system form a full-stack pitch for transferring human manipulation into reusable robot skills.

More
  • The SYNWorld model is intended to handle physical reasoning and action across tasks and robot bodies, while the OctoH-Hand supplies fine manipulation.
  • The OctoSense system combines a headband, electromyography wristband and hand exoskeleton to align human operations with a robot hand.
  • The company claims zero-shot cross-person generalization for electromyography data, a potentially important answer to the bottleneck of collecting robot-specific demonstrations.
  • The architecture treats data capture, world understanding, policy generation and real-world feedback as one loop rather than separate products.
  • The report is company-authored and offers no benchmark, deployment rate or failure analysis, so the technical claims remain directional.
Open at QbitAI Ask Claude ChatGPT

Space-computing startups face a four-week upload bottleneck

# Analysis GeekPark

Chinese space-computing startups say uploading an 800MB model to an orbiting satellite can take nearly four weeks, exposing the communications bottleneck behind the sector’s grand AI narrative.

More
  • The constraint is the uplink: legacy spacecraft links are designed for small commands, not large model files, and ground-station visibility arrives in short windows on each orbit.
  • Downlink capacity has improved because satellites send data to large ground receivers, while the reverse direction still lacks an operational high-throughput laser network.
  • The company Space Byte is reportedly pursuing “ground-to-space storage” before full computation, a pragmatic way to build the missing upload capability and a business case around it.
  • This is a valuable engineering reality check for orbital data centers: cheap energy and cooling do not matter if models, updates, and workloads cannot be moved to the spacecraft.
  • The sector map and startup count are interview claims, while the four-week figure is presented as a current practical estimate rather than a standardized benchmark.
Open at GeekPark Ask Claude ChatGPT

Researchers push post-training quantization toward one-bit models

# AI Leiphone

A researcher moving from ETH Zurich to Hong Kong Polytechnic University argues that extreme post-training quantization can fit large language models into much smaller devices without full retraining.

More
  • The reported motivation is a roughly 20-fold gap between model-size growth and hardware-memory growth, making brute-force hardware scaling insufficient.
  • The BiLLM approach isolates sensitive weight channels and compresses the rest to one bit, while SqueezeLLM uses dynamic two-bit allocation in groups of 128 weights.
  • The claimed benefit is lower memory traffic and power, not merely smaller disk files; that matters because bandwidth and memory movement often dominate edge inference.
  • The remaining barriers are activation and KV-cache quantization, hardware support for sub-two-bit arithmetic and hidden losses on complex reasoning tasks.
  • The researcher is Chinese and is moving to a Hong Kong university, giving the item a China research dimension even though the work was conducted in Europe.
Open at Leiphone Ask Claude ChatGPT

Personalized mRNA cancer vaccine shows positive Phase III melanoma result

# Signal Solidot

Chinese reports say Moderna and Merck's personalized mRNA vaccine, combined with Keytruda, extended recurrence-free survival after high-risk melanoma surgery; full data are still pending.

Also: IT Home

More
  • The trial enrolled more than 1,100 patients whose high-risk melanoma had been surgically removed, comparing the vaccine-plus-Keytruda regimen with Keytruda alone.
  • This is a therapeutic vaccine tailored to tumor-specific mutations, not a preventive vaccine for healthy people; it is designed to help the immune system recognize a patient's own cancer.
  • The strategic significance is that individualized mRNA oncology has reportedly crossed a Phase III efficacy milestone, moving beyond the technology's better-known infectious-disease use.
  • The supplied account gives no hazard ratios, absolute survival curves, or complete safety tables, so the size and durability of the benefit cannot yet be independently assessed.
Open at Solidot Ask Claude ChatGPT