China Desk

China News, Summarized 20 Oct 2026, 04:07 UTC 34 stories 22/23 sources

Today's brief

A growth beat lets Beijing spend the plenum run-up on discipline, not stimulus

Growth cleared; the Party talks about itself

Growth beat by a tenth. China's economy grew 4.5% in the third quarter, against a forecast of 4.4% and a three-year low of 4.3% in the second. September retail sales rose only 1.1%, and property investment is down 20.6% over nine months.

Two readings of one print. The National Bureau of Statistics says the economy "held up under pressure and moved steadily". CNBC sees one still leaning on factories and foreign buyers. Investinglive reads the beat as room to keep holding back big stimulus. That last reading is where my argument starts.

The plenum is about the Party. The Central Committee's fifth plenum runs 26 to 29 October and is devoted to Party self-governance, not the economy. On Monday the Party's mouthpiece ran an essay by Xie Chuntao, a vice-president of the Central Party School. He argues the Long March teaches members to uphold a strong "core", the Party's word for Xi.

Same message, two registers. Add the watchdog's no-way-out warning on Saturday. Add the investigation of Huang Qifan, the former Chongqing mayor, that I flagged on 17 October. The week reads as one argument: nobody sits outside the discipline.

Ding borrowed the vocabulary. At the Financial Street Forum, Ding Xiangqun, who runs the National Financial Regulatory Administration, promised to punish the key people behind major financial risk. She also promised to handle weak institutions through mergers and market exit, and to bring every financial activity under supervision. Named individuals, not just institutions: the anti-corruption grammar, applied to finance.

WHAT WE'RE WATCHING: whether the plenum communique on 29 October names expulsions from the Central Committee itself. That is where rhetoric either gets names or stays rhetoric.

Go deeper on this section: ClaudeChatGPT

Chen Deming's diagnosis is the EU's complaint

Weak demand, not overcapacity. Chen Deming, a former commerce minister and honorary chair of the Center for China and Globalization, a Beijing think tank, told its outbound-investment forum that China has "strong supply and weak demand", not overcapacity. His advice to companies was to go abroad.

Today's data agree with him. Fast Retailing, Uniqlo's parent, now earns more in North America plus Europe than in Greater China, for the first time. The region's share of its overseas revenue fell from 57% to 30% in five years, and it net-closed 51 mainland stores. Among China's ten biggest automakers, only Geely and Chery grew in January to September.

THE TELL: BYD executive vice-president Li Ke confirmed another 10 car carriers, a deal reported at over $1bn, taking its fleet from 8 ships to 18. One carmaker is spending on ships, not showrooms.

The catch. A home market that cannot absorb output pushes it abroad, which is Brussels's argument too. The Wire China's weekend essay argues that as long as China keeps its model, Europe will have to use more than import limits. Chen's diagnosis is the case against his conclusion.

Go deeper on this section: ClaudeChatGPT

Model-builders get millions; chroniclers get detained

A wage chronicler is detained. Feng Rui, who writes as Xiao Wu and edits the independent outlet Chiapas Dongfeng TV, was detained in Guangzhou on 14 September on a charge of "picking quarrels and provoking trouble". China Digital Times, a US site tracking Chinese censorship, says this is widely seen as retaliation. His team's annual review of labour incidents was censored and its WeChat account suspended.

Silence about the silence. Chinese editors archived seven articles on the case, and five have since been deleted, the site reports. One censored essay argued the charge should not be turned on people who document unpaid wages and layoffs.

Meanwhile, the model builders. Leiphone's headhunter sources say new graduates on foundation-model teams now start around 2 million yuan (~$282,000) a year. Top offers at ByteDance's Seed lab and Xiaohongshu, also called RedNote, reach 4 to 5 million yuan (~$563,000 to $704,000). Headcount is frozen at most big firms except model teams, DeepSeek and Qwen.

My read: a market prices the people who build the machine, while police handle the person counting everyone else's wages. The pairing is mine; neither source makes it.

Go deeper on this section: ClaudeChatGPT

Chinese open weights became the default base

Tencent built on Alibaba. Tencent's Hunyuan Frontier team lifts Alibaba's open Qwen3.8-27B from 0% to 19.84% on Terminal-Bench 4.0, level with GPT-5.6 Terra and Gemini 3.8 Flash. The method keeps only runs that pass full verification and has the model distil them into runbooks. It replays them in a clean sandbox, then fine-tunes. The paper concedes one model, three evaluations, no independent seeds.

A Silicon Valley startup did the same. Conway Research, a US startup eight months old, launched Underdog, an invite-only personal agent that runs entirely on the user's machine. Its Saluki 27B model derives from Qwen3.8-27B. Zhidx says it keeps 96% of full-precision performance at one-seventh the weight size. Backers include a16z and Anthology, the fund run by Anthropic and Menlo Ventures.

Google tested on GLM-5. Google Research's Harness², open-sourced under Apache 2.0, rewrites an agent's prompts, memory and tool guidance without touching weights. It was run on GLM-5, from the Chinese lab Zhipu, and on Gemini 3.5 Flash. It gained up to 11.5 points on WorkBuddy-Bench, a benchmark named after Tencent's agent product.

My read: nobody here asked Alibaba or Zhipu for permission, which is the point of open weights and what a licensing regime would change. They join Harvey, the US legal-AI startup that built its first model on Moonshot AI's Kimi K3.

Go deeper on this section: ClaudeChatGPT

Manus's agent has an ID, not a pass

Manus is back, on Beijing's terms. Manus, the agent startup that moved its base to Singapore in 2025, is independent again. The National Development and Reform Commission ordered Meta's roughly $2bn purchase unwound on 27 April, the first time a completed deal was reversed under China's foreign-investment security review. TMTPost, citing media reports, says it has raised over $500m at about $4bn.

Cue's ID card. Manus's new assistant Cue gives each agent its own email, phone number, wallet and computer. After a week, TMTPost found email works. Phone and payments need international accounts, and Cue cannot read WeChat or QQ, where Chinese users' schedules actually live. On a Q4 communications plan it was no better than ByteDance's Doubao or Tencent's WorkBuddy.

My read: this tests the 15 October claim that China's biggest apps decided an agent must be invited in. Alibaba's Qwen Go! borrows the user's own identity and can file Taobao price-protection claims. Cue brings its own identity and stays outside the gate.

Go deeper on this section: ClaudeChatGPT

Sixteen ships, four fronts around Taiwan

A choreographed Sunday. Taiwan's coastguard says it warded off 16 mainland vessels: 12 coastguard and research ships plus four fishing boats. They entered restricted waters at four places at once, including around Dongsha, which Taiwan controls. It called the operation "hybrid warfare". Experts quoted by SCMP read it as civilian ships being used to build a jurisdiction claim.

Not a one-off. Taiwanese official data show 244 mainland coastguard, research and government vessels around the island in July alone.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The slide found its owner. On 18 October, CITIC Securities called the AI-stock correction similar in size to new-energy's 2021–22 fall. TMTPost's third-quarter tally puts the damage at ChiNext down 27.8%. 210 of about 270 disclosed equity insurance-asset products lost money. China Life's first-half profit of 134.5bn yuan (~$18.9bn) drops to 57.6bn yuan (~$8.1bn) excluding market swings, so third-quarter marks may hurt.
  • DeepSeek's harness bet has a scoreboard. On 18 October I said DeepSeek now owns the whole agent loop. Arena now ranks V4.1-Flash-1017 the top open model on its WebDev board at 1,670, 12 points behind GPT-6 Sol (Max). Peak pricing is 2 yuan (~$0.28) per million input tokens and 8 yuan (~$1.13) per million output.
  • The memory squeeze reached a cheap phone. After the 20% and 22% shipment falls I tracked on 14 and 16 October, Honor raised the X80 Pro Max by 500 yuan (~$70), to 2,499 yuan (~$352). IT Home gives no reason; memory costs are my inference.

Go deeper on this section: ClaudeChatGPT

The river · 34

Chinese interconnect-chip maker reports 800,000 shipments and new 800G products

# QbitAI 17h

Zhongxing Micro says it has shipped 800,000 networking, storage and switching chips and unveiled products aimed at PCIe 6.0 and 800G AI clusters.

More
  • High-speed interconnect is a system bottleneck: adding accelerators does not deliver proportional throughput if nodes cannot exchange gradients, parameters and storage data efficiently.
  • The company claims its PCIe 5.0 switch reaches 55 GB/s and that a storage card outperforms a foreign comparator in a specified RAID test; these are vendor-reported results.
  • Its new PCIe 6.0 switch is rated for up to 144 lanes, while an 800G network interface targets large AI clusters and low-latency data movement.
  • The product range matters because Chinese compute suppliers need networking and storage components as well as processors; replacing a mature global stack requires long reliability testing and software qualification.
  • The report’s patriotic framing runs ahead of the evidence: shipment totals and demonstrations do not establish parity across every workload or customer deployment.
Open at QbitAI Ask Claude ChatGPT

China drafts technology export controls as the Fifth Plenum approaches

# Sinocism 5h

Chinese coverage foregrounds draft controls on AI and other technologies alongside Party governance and GDP updates, while highlighting uncertainty over what could ultimately be exported.

More
  • The Ministry of Commerce and Ministry of Science and Technology proposed adding technologies in AI, biomedicine and photovoltaics to China’s restricted or prohibited export catalogue.
  • The draft reportedly covers 23 AI technology categories, including large models, agents, chip design and computing clusters; public technologies may be excluded, while licensing rules remain unsettled.
  • The commentary says some registration requirements for covered exports, including transfers already underway, took effect immediately, despite the wider catalogue still being a draft.
  • The proposal offers facilitation for exports supporting cooperation among members of the planned World Artificial Intelligence Cooperation Organization, making the policy both a restriction and an industrial-diplomacy tool.
  • The same briefing reports official third-quarter GDP growth of 4.5% and Xi Jinping’s inspection of Inner Mongolia, while placing the upcoming Fifth Plenum in a Party-governance frame.
  • The final scope could change after public comments; the practical impact on model releases, software services and research collaboration remains unclear.
Open at Sinocism Ask Claude ChatGPT

Detention of labor-rights advocate Feng Rui raises fears of censorship

# China Digital Times 21h

Chinese labor-rights advocate and citizen journalist Feng Rui has been detained since September; supporters say prolonged questioning has harmed his health and followed publication of a detailed labor report.

More
  • Feng was reportedly detained in Guangzhou on a charge commonly translated as picking quarrels and provoking trouble, after years of documenting wage arrears, workplace accidents and labor disputes.
  • His team’s 2025 review covered 15 topics, including informal work, overwork, social security, automation, women’s work and disability, according to the report.
  • The outlet says the report was censored online, its associated account was suspended and Feng’s family faced pressure from authorities.
  • The case shows the sensitivity of documenting labor conditions independently, particularly when research aggregates disputes that are otherwise treated as isolated incidents.
  • The source reports concerns about his health and legal access but does not provide an official account of the investigation’s evidence.
  • The status of the report and its authors illustrates how censorship can restrict not only protest but also the creation of a public record about work.
Open at China Digital Times Ask Claude ChatGPT

Advocates renew calls to release US scholar Min Zin from Chinese detention

# China Digital Times 1h

A US citizen and Myanmar democracy advocate remains detained in China on alleged espionage grounds; supporters say they have little information about his condition or access to counsel.

More
  • Min Zin was detained in Kunming in June after traveling to China for an academic workshop, according to the source.
  • China’s foreign ministry said he was suspected of espionage that endangered national security; the US government has designated the case wrongful detention.
  • The advocates’ statement says there is little public information about evidence, legal access or his condition, making the case difficult to assess independently.
  • The case illustrates how national-security investigations can affect visiting scholars and cross-border academic exchange.
  • The account comes from advocacy groups and the China Digital Times report; the allegations against Min Zin have not been tested in a public trial in the information provided.
  • A durable resolution would require transparent legal process and access to counsel, not only diplomatic statements.
Open at China Digital Times Ask Claude ChatGPT

US and China seek concrete deals ahead of a Shenzhen leaders’ meeting

# SCMP 6h

Chinese reporting says both sides are racing to assemble investment and economic agreements before next month’s leaders’ meeting in Shenzhen, even as trade leverage remains contested.

More
  • Sources told the South China Morning Post that negotiations will continue in the coming weeks and still face obstacles to a complex package.
  • The discussions reportedly combine investment opportunities with institutional frameworks, making implementation and verification as important as headline commitments.
  • The negotiations unfold alongside mutual efforts to preserve leverage in extending a broad trade agreement.
  • The report relies on unnamed sources and does not identify finalized deals or their value.
Open at SCMP Ask Claude ChatGPT

Chinese maker prepares to sell a 0.1-nanometer electron microscope

# SCMP 7m

Hefei-based CIQTEK says its 200-kilovolt transmission electron microscope reaches a 0.1-nanometer information limit, a notable move into equipment dominated by US and Japanese suppliers.

More
  • The report gives a point resolution of 0.25 nm and an information limit of 0.1 nm, placing the instrument in the range of leading workhorse models.
  • Commercialization would matter beyond one instrument: electron microscopes are essential tools for examining advanced materials and semiconductor structures.
  • The available report is brief and does not establish independent benchmark results, production volume or customer adoption.
Open at SCMP Ask Claude ChatGPT

China’s pet-cat boom tracks urban time, housing and loneliness

# TMTPost 37m

China’s urban cat population reached 72.9 million in 2025, nearly 19.5 million more than dogs; the shift points to pets suited to smaller homes and less predictable schedules.

More
  • The cited industry white paper puts urban pet dogs at 53.4 million, while cat numbers have outpaced dogs since first overtaking them in 2022.
  • Cats are less dependent on daily walks and are easier to keep in rented or compact housing, a fit for a mobile population and smaller households.
  • The article argues that cat cafes, pet content and paid feeding services monetize companionship for people who cannot or do not want to own a pet.
  • Cat-market spending is growing faster than dog spending despite lower average spending per cat, suggesting a scale-driven consumer market rather than simply a shift to pricier pets.
Open at TMTPost Ask Claude ChatGPT

Huawei-Chery brand Zhijie moves toward its own vehicle production license

# TMTPost 37m

A Chinese regulatory filing proposes admitting Zhijie’s own company as a vehicle maker while Chery seeks to remove the same factory address from its license.

More
  • The move would shift formal manufacturing responsibility for Zhijie vehicles from Chery to a dedicated Zhijie legal entity, if the proposed admission is finalized.
  • China’s vehicle approval system makes the licensed manufacturer responsible for production consistency and product quality, so this is more than a branding change.
  • The structure described is an unusual division of roles: Chery retains control, Huawei leads operations, and Zhijie’s company could hold the production responsibility.
  • The factory remains at the same site; the report says the associated factory upgrades still list Chery as project owner, leaving asset and operating arrangements unresolved.
  • The change would make Zhijie’s corporate structure more distinct without severing its dependence on Chery’s manufacturing base.
Open at TMTPost Ask Claude ChatGPT

Tencent researchers train a terminal agent through repeated self-improvement

# AI Zhidx 18h

A Tencent-led research team proposes Terminal-RSI, alternating verified experience reuse and reinforcement learning to improve an agent’s command-line task execution.

More
  • The team uses Qwen3.8-27B as its base model and alternates recursive self-rewrite, which extracts reusable successful trajectories, with policy refinement through reinforcement learning.
  • Across four Terminal-Bench suites, the paper reports an average score increase from 20.68% to 50.58%; the result is a research evaluation, not evidence of general-purpose reliability.
  • The method filters and re-tests successful traces before using them, reducing the risk that an unverified run becomes a training example.
  • The researchers also adjust advantage estimation to avoid treating long environment logs as if they were additional meaningful agent decisions.
  • The paper’s comparison does not fully equalize the extra compute and supervision used to discover and rewrite trajectories, so the measured advantage is not a clean same-cost comparison.
  • The findings matter if verified experience can improve long-horizon tool use without excessive retraining or benchmark overfitting.
Open at Zhidx Ask Claude ChatGPT

Google’s Harness² uses execution evidence to improve agent scaffolding

# AI QbitAI 1d

Google Research’s open-source Harness² changes prompts, tools and memory through repeated trials; Chinese coverage emphasizes the possibility of improving agents without fine-tuning weights.

Also: InfoQ China

More
  • The framework proposes candidate changes to an agent harness, runs them against the same task and retains modifications supported by observable execution differences.
  • A second loop preserves improvement experience across tasks, aiming to avoid both fixed prompts and brittle task-specific tweaks.
  • Reported benchmark gains include up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on a legal benchmark, with added test-time compute as the tradeoff.
  • The framework is a research prototype, and results depend on the model, task and harness; they do not imply a universal agent upgrade.
Open at QbitAI Ask Claude ChatGPT

MiniMax launches a fast coding model to challenge DeepSeek on price

# AI Zhidx 13h

MiniMax released M3.1 Flash with a one-million-token context window and claims more than 200 tokens per second, putting it in direct competition with DeepSeek’s latest Flash model.

Also: Hacker News, V2EX

More
  • MiniMax lists prices of 1 yuan per million input tokens (about $0.14) and 4 yuan per million output tokens (about $0.56), with lower cache-read pricing.
  • The report cites benchmark strengths in long-running agents and cross-app automation, while DeepSeek leads on a software-engineering benchmark; the models appear to have different task profiles.
  • M3.1 Flash uses a sparse-attention design, according to the company, and the outlet reports a median output rate above 170 tokens per second.
  • The competitive signal is that Chinese labs are contesting not only model quality but also inference cost and latency in agent workloads.
Open at Zhidx Ask Claude ChatGPT

Chinese social-media post questions why foreign praise escapes scrutiny

# Culture China Digital Times 21h

A translated essay argues that Chinese online audiences and state media scrutinize foreign criticism of China while readily amplifying flattering visitor videos, revealing selective treatment of foreign influence.

More
  • The post contrasts suspicion of foreign involvement in negative coverage with enthusiasm for foreign influencers praising Chinese cities, food and infrastructure.
  • It cites official promotion of visiting creators and tourism campaigns, arguing that algorithmic visibility can make sponsored or curated praise look spontaneous.
  • The author’s critique is not that foreign visitors are necessarily insincere, but that audiences should ask who invited them, who paid and what ordinary experience is absent.
  • The essay shows that online arguments about foreign influence also concern domestic credibility: some users resent praise being used to dismiss their own complaints.
Open at China Digital Times Ask Claude ChatGPT

DeepSeek Flash tops a web-development leaderboard among open models

# AI Zhidx 17h

DeepSeek’s V4.1 Flash update ranked first among open models on Arena’s WebDev board, adding to evidence that Chinese coding models are competing on both capability and low inference cost.

Also: TMTPost

More
  • The article reports a score of 1,670, just 12 points behind the leading listed model, but leaderboard rankings can shift with usage and evaluation conditions.
  • DeepSeek’s API price varies by cache status and demand period; the outlet’s comparison says cached input can be especially cheap.
  • The source’s hands-on demos generated a 3D scene and a browser game for a small combined bill, illustrating inexpensive prototyping rather than production-grade reliability.
  • This is the same model release covered in other Chinese items in the batch, not a separate product event.
Open at Zhidx Ask Claude ChatGPT

Chinese startup launches a platform for computational medicine

# AI QbitAI 18h

Zheyuan Technology introduced WHIRLPO, a platform intended to support AI models of disease as a dynamic biological system rather than isolated molecules or targets.

More
  • The article argues that drug discovery needs models linking genes, cells, tissues and patient outcomes, because a molecule binding to a target does not establish safety or clinical efficacy.
  • The platform is described as infrastructure for training and running multiscale models, not as a validated disease predictor or drug candidate.
  • China’s latest pharmaceutical-industry planning reportedly lists computational medicine alongside AI, quantum computing and supercomputing as a priority technology area.
  • The central constraint is data: biological measurements are fragmented across modalities and institutions, with many high-value datasets requiring new experiments.
Open at QbitAI Ask Claude ChatGPT

China Telecom’s Xing model combines intent classification with parameter extraction

# AI Zhidx 15h

China Telecom’s AI lab open-sourced a 29B-parameter model that returns structured intents and linked fields, addressing a gap between deciding what a user wants and executing it.

More
  • The model outputs a constrained JSON structure intended to map natural-language requests onto entities, values and actions for downstream software.
  • The article reports under-150-millisecond end-to-end inference and strong zero-shot benchmark results, but those figures come from the outlet’s tests and require broader replication.
  • This design can reduce separate parsing stages in agents, but the model still needs safeguards for ambiguous instructions and actions that should not be taken.
Open at Zhidx Ask Claude ChatGPT

MemSilicon raises 400 million yuan and recruits a former Graphcore executive

# AI Leiphone 22h

MemSilicon raised more than 400 million yuan (about $56M) and hired former Graphcore executive Tao Lu to lead a new inference-chip effort alongside its RRAM business.

Also: GeekPark

More
  • The company plans to expand from resistive-memory commercialization into a first-generation LPU designed for large-model inference.
  • Lu’s stated target is fast decoding for full, unquantized models, with a system that balances memory capacity, compute and interconnect rather than maximizing one metric.
  • The article says the initial LPU will use more mature memory technologies; RRAM is not planned for the first chip.
  • MemSilicon’s technical proposition is to reduce the cost of repeatedly moving model weights and KV-cache data during token generation, a memory-system problem as much as a compute problem.
Open at Leiphone Ask Claude ChatGPT

US sanctions five Chinese entities over alleged trade-secret theft

# Policy SCMP 5h

Chinese coverage centers the US sanctions on firms and investors linked to a former Google engineer, framing the case as a new escalation in technology and IP enforcement.

More
  • The US State Department announced sanctions under the Protecting American Intellectual Property Act against five Chinese entities, according to the report.
  • Targets include Shanghai Zhisuan Technology, founded by former Google engineer Ding Linwei, venture firm MiraclePlus and three related entities.
  • Ding was convicted of seven trade-secret theft counts in January; the article says the new sanctions target entities suspected of involvement, not only the individual.
  • The case extends US enforcement from criminal prosecution into restrictions on organizations and financing networks, which may increase diligence risk for China-linked technology investors.
  • The report excerpt does not detail the sanctions’ precise scope or provide responses from the named entities.
Open at SCMP Ask Claude ChatGPT

China’s AI hiring boom widens pay gaps inside major tech firms

# Culture Leiphone 19h

Recruiters say Chinese giants are still paying exceptional sums for proven foundation-model talent even as hiring slows elsewhere, worsening internal pay inversion and labor-market stratification.

More
  • The report cites recruiter estimates that top foundation-model hires can command annual packages of 2 million yuan (about $280,000) or more, with especially high offers for scarce candidates.
  • Hiring is described as selective rather than broadly expansive: core-model teams compete for experienced researchers, while some application-side roles are contracting.
  • The salary gap between new AI specialists and longer-serving employees is presented as a recurring management problem, potentially weakening conventional seniority ladders.
  • Recruiters expect the market to divide further between top-tier talent with bargaining power and mid-career candidates facing more competition.
  • The figures are recruiter accounts, not comprehensive compensation data, and should not be treated as market averages.
Open at Leiphone Ask Claude ChatGPT

AliExpress plans to invest in AI hardware brands before they launch

# Analysis Leiphone 16h

Alibaba’s AliExpress and Tmall say they will combine early-stage investment with domestic and overseas retail support to develop more than 100 AI-hardware brands.

More
  • The plan moves the platform upstream from selling finished products to backing companies before their brands and products are established.
  • Alibaba can connect financing with product testing, marketing, fulfillment and overseas sales, giving it a stronger role in shaping which hardware companies reach customers.
  • For startups, the package may address gaps in global distribution and after-sales operations, but it also ties growth more closely to one platform’s channels.
  • The strategy reflects Chinese hardware firms’ turn toward overseas demand as domestic consumer-device competition and price pressure intensify.
Open at Leiphone Ask Claude ChatGPT

US agencies must remove DeepSeek and services built on its models

# Policy SCMP 6h

A US government notice bars federal civilian agencies from using DeepSeek products and related third-party services, citing national-security concerns.

More
  • The Department of Homeland Security notice reportedly covers products and services from DeepSeek and affiliated entities, including third-party systems developed using its models.
  • The order is a procurement and systems-control measure, not a general US ban on consumers or private companies using DeepSeek.
  • It signals that model provenance and downstream integrations are becoming part of federal software supply-chain screening.
  • The excerpt does not specify how agencies will identify indirect model use or how the policy applies to open-weight deployments.
Open at SCMP Ask Claude ChatGPT

Uniqlo’s growth is shifting from China toward Europe and North America

# Culture 36Kr 21h

Fast Retailing’s sales in Europe and North America have overtaken Greater China, while Uniqlo is closing weaker mainland stores and adapting more closely to local customers.

More
  • The article says Greater China’s share of overseas revenue fell from 57% five years ago to about 30%, as Europe and North America expanded faster.
  • Uniqlo closed 51 mainland China stores in fiscal 2026, leaving 851; mainland revenue reportedly grew only 3% in renminbi despite an 11.3% increase in Greater China on a yen basis.
  • The company’s China challenge is not simply fewer stores: management has criticized uniform, headquarters-led merchandising and called for more store-level decisions tailored to local consumers.
  • Europe now generates far more revenue per store than Greater China, and local product development and European managers are part of the growth strategy there.
  • The reversal is a useful indicator of China’s changing consumer market: international brands can remain present while losing their role as the main engine of global growth.
Open at 36Kr Ask Claude ChatGPT

China’s medical world-model push links hospitals, data and local computing

# Policy QbitAI 17h

A Nanjing conference showcased disease models, hospital data initiatives and a locally deployed medical-AI appliance, presenting world models as tools for simulating disease progression.

More
  • The event emphasized modeling patient states over time and comparing possible interventions, rather than only classifying images or predicting a static risk.
  • A consortium of hospitals announced plans for shared high-quality datasets, with patient data governed and kept within authorized environments.
  • A medical-computing appliance combines domestic hardware, operating systems and databases for hospital deployment, reflecting the policy emphasis on local control of sensitive data.
  • These announcements are early programs, not evidence of improved patient outcomes; clinical validation and external evaluation remain essential.
Open at QbitAI Ask Claude ChatGPT

A JD.com employee’s post turns a merchant dispute into a public fight

# Culture Leiphone 2h

A commentary says a JD.com sales employee reframed complaints about a merchant listing restriction as a rival-led smear, amid a separate legal clash between JD.com and Douyin.

More
  • The underlying dispute concerned a brand whose products were allegedly harder to find on JD.com after it used coupons on another platform.
  • The article argues that the employee’s public response shifted attention from platform treatment of the merchant to the motives of critics; this is the author’s interpretation.
  • The episode exposes tensions between platforms’ lowest-price positioning, merchant freedom to promote across channels and the role of public relations in Chinese e-commerce.
Open at Leiphone Ask Claude ChatGPT

A first-day sellout does not prove Apple’s foldable iPhone is a hit

# Culture GeekPark 2h

Chinese coverage stresses constrained launch supply and China’s cooling foldable-phone market, while the global rush also reflects pent-up demand for Apple’s first foldable.

More
  • The article says delivery estimates converged across several countries, which it interprets as evidence of centrally allocated limited supply rather than identical local demand.
  • It cites reports of low early production and assembly yield, but notes that these figures come from anonymous supply-chain sources and Apple has not confirmed them.
  • China’s foldable segment has been weakening, with Huawei a leading local player; limited Apple supply may encourage buyers to wait rather than switch immediately.
  • The high resale premiums and large preorder counts indicate interest, but without Apple’s first-batch volume they cannot establish total demand or sales potential.
Open at GeekPark Ask Claude ChatGPT

Agricultural drone pilots bring new work and new risks to rural China

# Culture TMTPost 53m

China’s agricultural-drone fleet is reshaping farm work, drawing younger operators back to rural jobs while leaving them exposed to weather, equipment failures and seasonal income swings.

More
  • The report cites 309,000 crop-protection drones in China in 2025 and 3.04 billion mu of annual operations, roughly 1.6 passes for every mu of national cropland.
  • Operators describe replacing days of manual spraying with hours of drone work, but also report crashes, plastic-bag damage and limitations in complex fields.
  • Some operators form teams and work split shifts during peak season; headline earnings can be high but are tied to intense, seasonal workloads.
  • The story places drone adoption alongside a sharp decline in agricultural employment and more than 15 million cumulative return-to-rural entrepreneurs, though the latter figure covers many kinds of business.
Open at TMTPost Ask Claude ChatGPT

Trump says he would meet Xi and Putin in China for nuclear talks

# Policy SCMP 7h

Donald Trump said he would accept a proposed three-way meeting with Xi Jinping and Vladimir Putin during his China visit, with nuclear weapons and denuclearization on the agenda.

More
  • Trump described the meeting as his suggestion and said it would depend on Xi’s willingness, according to the report.
  • A trilateral summit would place China at the center of nuclear diplomacy involving the two largest nuclear arsenals and Russia.
  • The statement is not evidence that Beijing or Moscow agreed to the meeting or its agenda.
Open at SCMP Ask Claude ChatGPT

Chinese 3D-printing startup targets small manufacturers instead of hobbyists

# Analysis Leiphone 21h

Lightweight Manufacturing is building lower-cost, high-throughput 3D printers for service farms and small businesses, rather than competing primarily for home users.

More
  • The founder argues that print farms are a sizeable but underserved customer segment where throughput, machine cost and downtime matter more than hobbyist features.
  • The company says its machines print four times faster than a named competitor and change colors five times faster than another; the claims are not independently tested in the report.
  • Its manufacturing partner is scaling production, suggesting the startup is positioning itself as an equipment supplier rather than only a consumer brand.
  • The commercial test is whether multi-head speed can coexist with print accuracy, stability and low operating costs at small-business scale.
Open at Leiphone Ask Claude ChatGPT

Chinese quantum-computing startup pitches a cloud path to usable hardware

# Analysis QbitAI 13h

A profile of a Chinese quantum-computing company argues that commercial access will require integrated hardware, software and cloud services rather than a headline qubit count.

More
  • The article describes quantum processors as potentially useful for selected molecular and optimization problems, but acknowledges that current systems remain noisy and limited.
  • Users face high hardware and operating costs, as well as a steep programming model based on quantum circuits rather than familiar application code.
  • The company aims to build a cloud platform at thousand-qubit scale, but qubit count alone says little about usable, error-corrected computation.
  • Claims that quantum machines can solve broad AI or scientific problems faster remain speculative until systems deliver reliable results on practical workloads.
Open at QbitAI Ask Claude ChatGPT

Chinese enterprises cite privacy and reliability as the real barriers to edge AI

# Analysis GeekPark 17h

A roundtable of Chinese software, chip and application vendors says customers want private deployment, quiet 24-hour operation and dependable response more than headline model size.

More
  • Participants describe elderly-care systems that must handle sensitive conversations and reminders locally, where noise, uptime and privacy are adoption requirements.
  • Enterprise customers are said to reject cloud processing for financial, order and other core records, especially in state-owned and regulated organizations.
  • The speakers argue that edge AI succeeds only when hardware, software support and service delivery fit a specific job; deploying a model on a device is not enough.
  • These are vendor and partner perspectives, but they reveal how data-control expectations shape China’s enterprise AI market.
Open at GeekPark Ask Claude ChatGPT

Cell Press investigates bribery allegations against a journal editor

# Culture SCMP 3h

Cell Press says it is investigating after Peking University biologist Rao Yi accused a leading medical-journal editor of soliciting payments from mainland authors.

More
  • The allegation concerns the integrity of editorial decisions and potential conflicts of interest in international academic publishing.
  • The report provides no conclusion from the investigation and does not establish that the editor accepted bribes.
  • The episode highlights wider concerns about opaque publication practices and the pressure to publish in high-prestige journals.
Open at SCMP Ask Claude ChatGPT

Fudan researcher says embodied AI’s bottleneck is not a new model architecture

# Analysis Leiphone 1d

An interview with Fudan researcher Jia Xiaosong argues that embodied systems need better closed-loop evaluation and data, not just successive VLA or world-model architectures.

More
  • His earlier driving research used latent-world simulation to reduce the interaction cost of training reinforcement-learning policies.
  • He says modular perception, prediction and decision pipelines can lose information between stages and optimize metrics that do not improve final driving behavior.
  • End-to-end systems can align modules around the task outcome, but intermediate signals may still help with supervision and interpretability.
  • The researcher’s shift from autonomous driving to dexterous hands and egocentric data reflects how similar evaluation and interaction problems are reappearing in robotics.
Open at Leiphone Ask Claude ChatGPT

DeepSeek’s robotics investments look beyond its foundation models

# Analysis Jiemian 21m

An analysis argues that DeepSeek is using investments in robotics and embodied-AI startups to build relationships with hardware, talent and testing resources while keeping its main research focus on foundation models.

More
  • The article links DeepSeek’s investment activity to a staged strategy: strengthen models first, while gaining early access to physical-AI capabilities through partners.
  • It describes Unitree as a potential source of robot hardware and motion-control experience, and other startups as exploring perception and embodied models.
  • That strategic interpretation is the author’s thesis; the article does not show that DeepSeek has committed to building a robot business itself.
Open at Jiemian Ask Claude ChatGPT

China’s Double 11 becomes a test of whether shopping agents can retain users

# Culture TMTPost 53m

Chinese platforms are using shopping coupons and agent features to turn Double 11 traffic into repeat AI use, but the available figures show trials more clearly than durable purchasing habits.

More
  • Douyin and Alibaba are attaching promotions and assistant features to shopping, while Alibaba’s Qwen Go is pitching agents for tasks such as refunds and price protection.
  • The article cites large cumulative counts of users who have tried AI shopping, but says publicly available data do not show how many return or how often they transact.
  • For merchants, the real test is incremental sales or lower service costs, not orders merely routed through an AI interface.
  • Agents that rank products also change which sellers get discovered, creating tension between consumer convenience and platforms’ advertising and merchant economics.
Open at TMTPost Ask Claude ChatGPT

China’s photonic quantum-computing industry faces an engineering test

# Analysis Leiphone 21h

A Chinese quantum-computing company argues that commercial progress depends on stable, manufacturable systems and lower optical loss, not only laboratory claims of quantum advantage.

More
  • The company says its latest platform uses thin-film lithium-niobate photonic chips and reports low propagation and coupling losses as key scaling measures.
  • Photonic systems can operate without the deep cryogenic environment required by some other quantum approaches, potentially easing data-center integration.
  • The article compares several national and technical approaches, but many performance claims concern specialized sampling tasks rather than useful commercial workloads.
  • Reliable logical qubits, software maturity and customer value remain unresolved; a claimed quantum advantage is not equivalent to a practical industrial advantage.
Open at Leiphone Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 1/20
Pekingnology 200 0/20
Ginger River Review 200 0/20
Sinocism 200 1/20
QbitAI 200 10/10
36Kr 200 7/30
Solidot 200 9/18
Leiphone 200 14/20
GeekPark 200 4/30
IT Home 200 12/60 +41 over cap
SCMP 200 10/50 +12 over cap
Initium 200 4/15
Jiemian 200 8/30 +10 over cap
China Digital Times 200 3/15
The Wire China 200 0/50
Techmeme 200 1/15
Hacker News 200 1/20
V2EX 200 8/40 +30 over cap
Zhidx 200 9/20
InfoQ China 200 8/20 +8 over cap
New Wisdom 200 8/15 +3 over cap
TMTPost 200 8/19 +10 over cap