China Desk

China News, Summarized 29 Aug 2026 34 stories archived day

Today's brief

The Pentagon's China list is now being answered in court, not by lobbyists

CXMT takes Section 1260H to a Washington judge

CXMT sued the Pentagon on Friday. ChangXin Memory Technologies — the DRAM maker whose first-half revenue of 150.3bn yuan (~$21bn) made it, by Bloomberg's count, China's most valuable listed business — filed in the US District Court for the District of Columbia to get off the Section 1260H list of firms the Defense Department says support China's military. Its chips are civilian, it says. Reuters has the filing.

The procedural history is the argument. Designated in January 2025, CXMT says it spent over a year submitting information; the Pentagon published a notice in February 2026 saying it would be removed, withdrew that notice the same day, then relisted the company in June without explaining the reversal. The complaint calls the decisions "arbitrary." Defense declines to comment on live litigation.

This is now a standard play. Xiaomi sued and got itself delisted in 2021; YMTC filed in December 2025, WuXi AppTec and Alibaba in June. Chinese champions have decided the US Administrative Procedure Act is a usable instrument — and CXMT chose to file three and a half weeks before Xi is expected in Washington on 24 September.

Then Friday's follow button paid off. I noted that CXMT's new Weibo account followed Xiaomi's phone division first. Saturday, the two confirmed that CXMT's LPDDR6 will debut in the Xiaomi 18 Fold next month alongside Xiaomi's in-house 3nm Xring O3. Domestic memory, domestic silicon, one flagship, one lawsuit.

And memory has passed the processor. A hardware leaker relayed by GeekPark puts a 12+256GB memory package at roughly 2,200 yuan (~$310) to procure against about 500 yuan (~$70) a year ago, while the newest flagship SoC went from ~1,500 to 2,000 yuan (~$282). Trade chatter, but the ordering is the point: memory is now the largest line in a phone's bill of materials.

Go deeper on this section: ClaudeChatGPT

Xi lands in Bishkek before he lands in Washington

Xi meets Putin on Monday. The Kremlin says Power of Siberia 2 will come up on the sidelines of the Shanghai Cooperation Organisation summit in the Kyrgyz capital — their second meeting this year. Foreign-policy aide Yury Ushakov said the issue "will be discussed bilaterally," per SCMP. The 1,600-mile line would move up to 50bn cubic metres a year via Mongolia.

The tell is who keeps announcing it. A memorandum exists; a price never has. Moscow puts the pipeline on the public agenda, Beijing declines to say what it will pay, and Chinese readouts of these meetings lead on multipolarity rather than gas. That pattern survived Putin's Beijing visit in May intact.

Treasury kept firing anyway. The US sanctioned a Hong Kong company Friday for allegedly laundering funds for a designated Iranian exchange house, the latest of several Chinese entities hit in days, with Beijing promising retaliation, SCMP reports. Nobody is protecting the atmosphere before 24 September.

Quieter statecraft: AI is now written into the family-doctor contract. A health-commission notice known as document 273 pulls hospital specialists into primary-care contracting and widens the disease list to include infectious disease by end-2027, then tells providers to "deeply advance" AI in those services — analysed in TMTPost. Follow-up volume rises; grassroots headcount does not.

The vendors are already sized for it. iFlytek's medical assistant covers 828 counties and over 77,000 primary institutions, claims above 80% share where AI triage is deployed, and grew that revenue 47.5% in the half. In Tianjin, WeDoctor reports one health manager covering 2,600 patients, about 3.5 times conventional throughput.

Go deeper on this section: ClaudeChatGPT

Alibaba went sparse, Meitu went to the bank

Qwen3.8-Flash shipped Wednesday night, open-weight as Qwen3.8-Flash-Next: a multimodal mixture-of-experts model with 125B main parameters plus 51B of N-gram embedding, only 6B active per token, at 0.8 yuan (~$0.11) per million input tokens and 2.7 (~$0.38) out.

The attention stack is the interesting part. Three of every four layers use gated delta networks compressing history into fixed-size state; the fourth keeps global attention for exact retrieval, with a compressed indexer selecting micro-blocks rather than re-reading context. Alibaba claims up to 7.6x prefill and 4.9x decode kernel speedups at 1M tokens, and in its Qwen office agent, 75% less token consumption per task — InfoQ has both.

Meitu is the one actually banking AI money. The 18-year-old photo-editing company reported first-half revenue of 2.21bn yuan (~$311m), up 22.1%, adjusted net profit 650m yuan (~$92m), up 39.5%, on 18.44m paying subscribers. Its productivity apps carry ARR of about 620m yuan (~$87m), and compute-credit spending roughly doubled across the half.

Set that against ByteDance's Doubao. A Jiemian commentary puts Doubao's daily revenue under 1m yuan (~$141,000) against 200m-plus daily users, and cites trackers showing 18.65m monthly users lost from May to July after a paid tier arrived. Selling a workflow tax beats selling a platform, for now.

Go deeper on this section: ClaudeChatGPT

A robot made 55-step ice cream in Shanghai

Sharpa put a two-armed robot behind a DQ counter. At the chain's Wujiang Road store in Shanghai it takes a metal holder from a steriliser, pulls a paper cup, draws soft serve, adds three scoops of Oreo, blends, and inverts the cup without it falling — the upside-down test the brand has used since 1940. TMTPost watched the cycle.

The specifics make it a product, not a demo. 55 steps, about 90% success, touch feedback in 98% of them, and zero modification to a store that imports every fixture. Chief executive Li Yifan, also a co-founder of lidar maker Hesai, calls purpose-built robot environments "phone booths" and says doing 54 of 55 steps creates a job rather than filling one. Sharpa has raised over 4.5bn yuan (~$634m) at above 20bn yuan (~$2.8bn).

Now read the Beijing robot games two ways. Solidot records Tiangong, from the Beijing Humanoid Robot Innovation Centre, at 8.64 seconds over 100 metres against Bolt's 9.58 — then spends the piece on why it isn't sprinting: nearly a second to react against Bolt's 0.146, more than 50 steps against his 41, no deceleration, and a stop achieved by hitting a crash mat.

AgiBot's version is a medal table. New Wisdom relays 18 golds and 46 medals, including seven of eight dexterous-hand events and six of twelve scenario events — library shelving, hotel rooms, fire response — on production units. Western wires led on the record and used the crash mats as colour; CBS mentioned the stretchers. Correction: I gave Tiangong 8.86 on 25 August. The scenario events are the real scoreboard.

Go deeper on this section: ClaudeChatGPT

A programmer died in a company toilet

Shenzhen refused to call it a workplace injury. A 39-year-old programmer parked in his office building's second basement at 8:56am on 23 April, asked a guard where the second-floor toilet was at 9:00, went in a minute later, and was found unconscious there at 11:28. He was declared dead at 12:50. Solidot has the file.

The stated reason is the story. The city's human-resources bureau ruled on 3 July that his desk was on the 11th floor, that he had clocked in and gone to the second-floor toilet without going to it, and that he was doing nothing work-related. Recognition is worth 1,130,040 yuan (~$159,000) plus funeral and dependants' payments. The family is appealing. Two and a half hours passed before anyone looked.

A second argument is running between two cities. In July, Hong Kong police searched two independent bookstores on suspicion of sedition and arrested five people; one clerk was led away in a T-shirt reading "I am a bookstore clerk." Initium's podcast traces the slogan to Taiwanese bookstores, concert stages and a post by Taiwan's president — then to a fight over whether that was solidarity or consumption.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • MiniMax: the ARR figure I quoted twice is the softest number in the report. TMTPost's read of the interim: recognised first-half revenue was $116.6m against the $800m annualised claim, which Huxiu suspects is one week times 52 and unaudited. Gross margin came in at 17.9% against Morgan Stanley's 29.5%, because enterprise volume was bought with price cuts — M3 doubled API prices on 1 June, the stock fell 15.71% that day, permanent half-price followed within a week. Adjusted net loss $293m, up 111.2%.
  • Gyirong passed 72 hours; Yunnan started flooding. Excavators are reopening the route to the Gyirong border checkpoint with the barrier lake still a breach risk, and CCTV named the officer, Huang Peng, 34, filmed directing people away minutes before the mudslide. Meanwhile Mohan, on the Laos border, took 187.5mm (7.4 inches) of rain in 24 hours; 2,226 people moved, provincial response at level three, 20,500 relief items plus 10,000 from central stores.
  • The discipline machine went back to county work. A day after a legislative committee stripped the army's two most senior officers of their seats, the same apparatus expelled a former vice mayor of Tongliao in Inner Mongolia for, among other things, falsifying shantytown-redevelopment figures — a political-performance charge, not merely a bribery one — plus a former deputy chair of Ulanqab's legislature, with a Chongqing Airport Group executive newly under investigation.
  • Biren's first post-IPO half is 20x revenue and 23% cash. Revenue 1.236bn yuan (~$174m), up 1,997.6% and more than all of 2025; loss narrowed 76% to 380m yuan (~$54m); gross margin 42.7%, down from 53.8% last year as cluster deals drag in third-party hardware. Receivables 1.117bn yuan (~$157m), collected at 23.3%; R&D exceeds gross profit. AIX notes its HK$101.7bn (~$13bn) cap is a third of A-share peers with the same losses.

Go deeper on this section: ClaudeChatGPT

The river · 34

Inference software can change a model's answer even with identical weights

# QbitAI

A long-context test found that attention kernels, KV-cache precision, quantization, and tensor parallelism changed token choices and broke tool calls on the same GPU and weights.

More
  • The experiment captured logits across more than 100,000 tokens from a Qwen model and found that switching attention backends eventually changed the top-ranked token, despite identical prompts, hardware, weights, and forced histories.
  • BF16 KV cache stayed stable, while INT4 produced rapidly rising top-token flips and failed tool calls; INT8 sometimes recovered, showing why short-context spot checks can miss production failures.
  • The causal mechanism is numerical drift from floating-point precision, accumulation order, and CUDA kernel behavior, amplified through autoregressive generation rather than a semantic difference in the model weights.
  • The test found that a community INT8 weight-quantization path preserved activation precision better than some official FP8 and FP4 paths, while an FP4 path approached a 50 percent flip rate near 88,000 tokens in that setup.
  • A striking tensor-parallel result was that one, two, and four GPU configurations did not behave monotonically: single- and four-GPU runs completed a tool call while the two-GPU run failed.
  • The source is a forum experiment, not a peer-reviewed benchmark, but it exposes a real systems boundary: inference reproducibility belongs to the full kernel, cache, quantization, and parallelism stack.
  • For agent operators, long-context regression tests should include exact tool-call trajectories and numerical stability, not just answer quality on short prompts.
Open at QbitAI Ask Claude ChatGPT

A returned scholar becomes a Tsinghua AI class adviser

# QbitAI

Xu Mengdi returned from Stanford to Tsinghua in 2025 and now advises its elite AI cohort, bringing a robotics view that emphasizes adaptation over simply scaling models and data.

More
  • Xu studied vehicle engineering before moving through robotics and machine learning at Johns Hopkins and Carnegie Mellon, then working with Toyota Research, MIT-IBM, Google DeepMind, and Stanford labs.
  • Her RoboTool work lets robots choose tools and compose executable plans from goals, scenes, and capabilities, including unconventional uses such as turning objects into bridges or hooks.
  • Her central research claim is that a robot cannot encounter every future situation in its training set, so generality requires fast adaptation from a small number of demonstrations.
  • Prompting Decision Transformer applies an in-context-learning idea to reinforcement learning by using short task trajectories to infer a new goal without updating model parameters.
  • The appointment is a talent signal: elite Chinese AI education is recruiting researchers who connect foundation-model ideas to physical control, safety, and real-world deployment.
  • It also shows the value of return migration beyond headcount, as overseas research experience is being folded into the training of China's next generation of AI engineers.
Open at QbitAI Ask Claude ChatGPT

A Chinese coastguard war game puts weapons use at 0.3 percent

# SCMP

A Chinese coastguard simulation estimates a very low chance of firing in a disputed-water encounter, revealing how researchers model escalation, ambiguity, and restraint.

More
  • The scenarios involve fishing boats or small fast craft probing a patrol ship, forcing operators to distinguish ordinary activity from coordinated attempts to expose a weakness.
  • The reported 0.3 percent weapons-use probability is a model output, not a forecast or evidence that real-world confrontations are safe.
  • The technical value is in the decision problem: Chinese researchers are treating maritime escalation as a classification and control problem under incomplete information, not simply a question of force superiority.
  • The exercise highlights a preference for calibrated coercion, where coastguard presence and maneuvering achieve pressure while avoiding the political cost of a shot.
  • The number should be read alongside assumptions about actor behavior, rules of engagement, sensor confidence, and repeated encounters; changing those inputs could radically alter the result.
  • The broader signal is that China is institutionalizing gray-zone scenarios and quantifying escalation risk around disputed waters.
Open at SCMP Ask Claude ChatGPT

China's Middle East strategy faces a test as the Iran war drags on

# SCMP

Six months into the US-Israeli war with Iran, Gulf states are building more self-reliant defense networks, challenging China's strategy of regional involvement without firm security commitments.

More
  • The report says Gulf governments increasingly doubt that either Washington or Beijing can protect them from attacks, a judgment that limits China's leverage even as its economic role grows.
  • China's approach has emphasized diplomacy, trade, and political relationships without extending treaty-style military guarantees; the conflict exposes the gap between influence and responsibility.
  • A regional shift toward autonomous defense procurement and partnerships could reduce the value of Chinese diplomatic balancing while increasing pressure on Beijing to clarify its red lines.
  • The issue is not simply whether China can mediate: it is whether Gulf states will continue treating China as a strategic alternative if security still depends on other networks.
  • The story places China's Middle East policy at a crossroads between avoiding entanglement and accepting the costs required to become a credible security actor.
Open at SCMP Ask Claude ChatGPT

CXMT sues the Pentagon over its Chinese military-company designation

# Jiemian

Chinese coverage stresses CXMT's civilian-chip argument, while US reporting frames the case as a challenge to its Pentagon blacklist designation and alleged military ties.

More
  • CXMT filed in a US federal court seeking removal from a list that identifies companies alleged to support China's military, according to the supplied reports.
  • The company says its memory chips serve civilian and commercial uses and that the designation has harmed its reputation and business interests.
  • The dispute shows how entity lists can impose commercial costs before a technical or legal finding on the underlying relationship is resolved.
  • Because CXMT is China's leading memory producer, the case also tests how far Chinese semiconductor firms will use US courts to contest national-security restrictions.
  • A ruling or settlement would clarify the evidentiary standard and could influence future challenges by Chinese chip companies.
  • The company-side claims and the Pentagon's decision not to comment on pending litigation should not be treated as a final finding.
  • The technology angle is strategic rather than a new memory milestone: legal status now directly affects access, customers, and financing.
Open at Jiemian Ask Claude ChatGPT

After two decades of rural expansion, China is putting counties first

# TMTPost

China's new commerce policy shifts emphasis from extending goods to every village toward making county seats efficient hubs for services, consumption, and regional logistics.

More
  • The earlier phase used village stores, appliance subsidies, e-commerce, mobile payments, and parcel delivery to solve access and affordability for physical goods.
  • The supplied figures show rural online retail exceeding 3 trillion yuan, roughly $423B, in 2025, while parcels now reach almost every administrative village.
  • The next constraint is service density: restaurants, clinics, childcare, cinemas, and elder care must keep local staff and facilities busy, unlike a refrigerator that can be shipped from a distant factory.
  • County seats can aggregate enough demand to support services while remaining geographically closer to rural residents than major cities.
  • The policy therefore marks a transition from coverage to network design, with counties acting as the middle layer between national supply chains and dispersed households.
  • It also reflects changing consumption: rural service spending reached 8,173 yuan per person in 2025 and represented 40.3% of consumption, according to the supplied analysis.
  • The important test is whether county investment creates viable service ecosystems or merely moves underused commercial capacity one level upward.
Open at TMTPost Ask Claude ChatGPT

MiniMax grows fast while discounts and compute deepen its losses

# TMTPost

MiniMax reported first-half revenue of $116.6 million, up 283.1 percent, but adjusted losses widened to about $293 million as enterprise growth relied on lower pricing and heavier research spending.

More
  • The reported IFRS loss of $358 million improved mainly because a pre-IPO preferred-share revaluation charge disappeared after listing; operating losses instead rose to about $327 million.
  • Research spending reached $297 million, roughly 2.5 times revenue, while adjusted net loss exceeded the full-year 2025 figure of $251 million.
  • Enterprise and open-platform revenue rose to $73.93 million, about 63.4 percent of total revenue, shifting the company from consumer applications toward selling model capability into business workflows.
  • The article says gross margin fell to 17.9 percent even as enterprise mix increased, implying that volume was purchased through discounts rather than pricing power.
  • MiniMax's M3 model reportedly cut per-token computation through sparse attention and reached a 1 million-token context, but an attempted price increase was later reversed to roughly the old level.
  • Its H3 video model generated strong downloads, yet the financial question remains whether video, enterprise APIs and consumer products can cover training and inference costs without continued financing.
  • The strategic read is not simply growth versus loss: MiniMax is proving demand for Chinese models while testing whether that demand can support durable margins in a market crowded by DeepSeek, Kimi and other low-cost rivals.
Open at TMTPost Ask Claude ChatGPT

Biren's revenue surges as Chinese GPU demand reaches real orders

# TMTPost

Biren reported first-half revenue of 1.236 billion yuan, roughly $174 million, up nearly 20 times, showing that Chinese AI-chip customers are beginning to buy integrated systems rather than just fund benchmark narratives.

More
  • About 1.168 billion yuan, roughly $165 million, came from bundled computing solutions spanning chips, boards, servers, software and cluster integration.
  • Biren says it passed supplier certification at major internet companies and began volume delivery, while a northern China cluster reached several thousand cards.
  • Its BIRENSUPA software stack supports domestic models including DeepSeek, Qwen and GLM, making deployment compatibility part of the product rather than an afterthought.
  • The company reported a 42.7 percent gross margin and an adjusted loss of about 337 million yuan, roughly $47 million, but research spending reached 804 million yuan, roughly $113 million.
  • The near-20-fold growth is amplified by a low base: first-half 2025 revenue was only 59 million yuan, roughly $8 million, and the business is still concentrated in large solution deliveries.
  • The main risks are not only chip performance but sustained yields, software maturity, margins and whether internet customers continue buying beyond initial substitution orders.
  • Biren's results are an important marker for China's semiconductor strategy: export constraints and domestic demand are converting national substitution from a financing story into a procurement test.
Open at TMTPost Ask Claude ChatGPT

A Shanghai ice-cream shop may be a real test for humanoid robotics

# TMTPost

Sharpa's two-arm robot makes a Dairy Queen Blizzard through a 55-step workflow with no store redesign, offering a more credible test of embodied generalization than a trade-show demo.

More
  • The robot handles cups, dispensing, scooping, mixing and the brand's inverted-cup quality check inside an operating store.
  • Sharpa says the system reaches about 90 percent success and relies on tactile control and force sensing, not only a scripted pick-and-place sequence.
  • The company deliberately chose Dairy Queen's rigid imported equipment and materials so it could not simplify the environment around the robot.
  • Its longer-term thesis is that restaurant, hotel and retail tasks resemble household manipulation, creating data and reusable atomic skills that could eventually transfer into homes.
  • The reported founders come from Hesai, and the company has raised more than 4.5 billion yuan, roughly $630 million, at a valuation above 20 billion yuan, roughly $2.8 billion, according to the source.
  • The key caveat is commercial proof: one store and a 90 percent success rate do not yet establish labor savings, maintenance economics or reliable deployment across locations.
  • The story matters because it shifts the embodied-AI test from impressive movement to unmodified environments, long-horizon manipulation and measurable operational value.
Open at TMTPost Ask Claude ChatGPT

Kong Shaoxun appointed Gansu's deputy Party chief and government chief

# Jiemian

A brief local-government notice confirms Kong Shaoxun has taken Gansu's two key provincial leadership posts, a meaningful personnel change near the top of a province.

More
  • The source identifies him as deputy secretary of the provincial Party committee and secretary of the provincial government Party group.
  • That combination normally places the official in line to lead day-to-day provincial administration, though the report does not state whether a formal governor appointment has followed.
  • The notice's emphasis on studying Xi Jinping's inspection instructions reflects the standard Party-government chain of political accountability.
  • The move matters less for a policy announcement than for reading the provincial succession and management structure.
Open at Jiemian Ask Claude ChatGPT

China reports 2.45 million PFLOPS of installed AI capacity

# IT Home

China's data authority says national AI computing capacity reached 2.45 million PFLOPS by July, with more than 85 percent concentrated in major national hubs.

Also: Jiemian

More
  • The figure is reported in FP16 and describes aggregate installed capacity, not necessarily the amount being productively used.
  • The eight national computing hubs and three power-computing coordination zones account for more than 85 percent of the reported total.
  • Only 1.45 million PFLOPS is currently covered by the national monitoring and scheduling platform, leaving a meaningful gap between capacity and centralized visibility.
  • The policy direction is clear: Beijing wants computing resources treated as coordinated infrastructure, with power supply, scheduling and utilization monitored together.
  • The number matters for model builders only if utilization, interconnect quality, memory and energy availability keep pace; PFLOPS alone cannot establish effective training throughput.
  • The same announcement says China authorized more than 3,400 agent-related patents in 2025, more than twice the prior year's growth rate, suggesting policy attention is moving from raw capacity toward applications.
Open at IT Home Ask Claude ChatGPT

Anthropic's automated researcher improves AI safety at machine speed

# AI QbitAI

Anthropic reports that Claude searched papers, designed training, and improved ten alignment tasks at roughly $4 an hour, while acknowledging that humans still define the goals and tests.

More
  • The automated alignment researcher closes a loop from literature search to hypothesis, data generation, fine-tuning, and evaluation, with each 30-minute experiment informing the next.
  • On a deception task, the best automated method reportedly closed 82 percent of the safety gap and averaged 85 percent across runs, compared with 20 percent for six human researchers given one attempt each.
  • The comparison is not fully fair: Claude could run repeated experiments and learn from outcomes, while human experts submitted a single method, so the result measures iterative search as much as research insight.
  • A weaker Sonnet model reportedly improved an early Opus model, closing about 65 percent of the selected gap with more than 2,000 examples and approaching the production model's 72 percent.
  • The reported API cost was about $4 per hour versus $150 per hour for human researchers, but the cheap loop depends on access to compute, a target model, evaluation infrastructure, and human-defined boundaries.
  • Anthropic's own caveat is the important one: the system optimizes what the benchmark specifies, and a more compliant model can still fail on risks or capabilities outside the test suite.
Open at QbitAI Ask Claude ChatGPT

Tencent releases a 770-billion-parameter Hunyuan preview model

# AI GeekPark

Tencent open-sourced Hunyuan Hy4 preview with 770 billion total parameters, a 1-million-token context window, and low API prices aimed at office, coding, and research workloads.

More
  • Hy4 expands from Hy3's 295 billion total parameters and 21 billion activated parameters to 770 billion and 49 billion, while quadrupling the advertised context window from 256K to 1M tokens.
  • Tencent prices input at 6 yuan per million tokens (about $0.85), output at 18 yuan (roughly $2.54), and cached input at 0.3 yuan (about $0.04), positioning it as a high-volume enterprise utility.
  • The company says 163 experts blind-tested 203 engineering tasks and gave Hy4 an average 2.99 out of 4, narrowly ahead of two other Chinese models; these are internal evaluation results, not an open benchmark.
  • Tencent also says the model contributed to optimizing its own training methods, data strategy, evaluation, and low-level operators, improving inference throughput 31.8 percent against a baseline.
  • The real product question is long-horizon execution: reading many documents, calling tools, recovering from errors, and delivering usable software rather than producing attractive one-shot text.
  • Hy4's deployment across Tencent's office, coding, content, and assistant products shows a vertically integrated Chinese model strategy in which distribution can matter as much as benchmark rank.
Open at GeekPark Ask Claude ChatGPT

OpenClaw's fad is giving way to agent harnesses

# Analysis QbitAI

A Chinese analysis argues that OpenClaw has not died but has been absorbed into vendor products, as developers move from assembling personal agents to using integrated runtimes and harnesses.

More
  • The article says OpenClaw reached more than 380,000 GitHub stars and spawned over 30 domestic derivatives, including products from major Chinese internet companies.
  • Its original appeal was permission-rich local autonomy: connecting email, browsers, terminals, calendars, and chat tools to a model that could run continuously without supervision.
  • The newer competition packages context management, file operations, command execution, sandboxing, long tasks, and tool orchestration into first-party agent runtimes such as Claude Code and Codex.
  • This is a useful product-cycle observation: open-source projects can define an interface and market before incumbents absorb the underlying workflow into more reliable, managed systems.
  • Chinese coverage treats the shift as a move from novelty and installation services toward infrastructure and harness quality, where stateful context, permissions, and recovery determine daily usefulness.
  • The question is whether OpenClaw retains a community advantage in extensibility, or whether vendors win by making the same capabilities safer and less labor-intensive.
Open at QbitAI Ask Claude ChatGPT

Alibaba's Qoder moves coding agents from IDEs to desktop workflows

# AI QbitAI

Alibaba says Qoder can turn natural-language requests into tested applications and modify existing repositories, using a plan-code-run-verify loop rather than merely generating snippets.

More
  • The agent reads an existing repository, infers its architecture, produces a change plan, edits multiple files, runs the project in a real environment, and revises when the result fails.
  • Its computer-use and browser-use layers matter more than the desktop mascot: they let the model verify generated software against actual interfaces instead of stopping at syntactically plausible code.
  • The company says Qoder has served 6 million users and 100,000 businesses, and supports more than 40 connectors, 70 plugins, and 20,000 skills; these are company-reported adoption figures.
  • The product reflects China's push to make agentic coding a general business tool, not just an IDE assistant, by connecting repositories to project management, cloud services, and internal systems.
  • The unresolved risk is governance: automatic approval and cross-system execution can turn a small coding error into a production change unless permissions, sandboxes, and review gates are explicit.
Open at QbitAI Ask Claude ChatGPT

China sends nine tunnel rescuers to Nepal flood zone

# Policy SCMP

China sent nine tunnel specialists to Nepal after a direct request from Kathmandu, as rescue operations faced hundreds missing on the Chinese side and thousands in Nepal.

Also: Jiemian

More
  • The source reports seven confirmed deaths and 544 missing on the Chinese side, with at least 675 dead and nearly 2,500 missing in Nepal.
  • The deployment is specialized rather than symbolic: tunnel experts are being used where collapsed infrastructure and difficult terrain complicate search operations.
  • China's assistance extends its domestic emergency-management capabilities across the border and gives Kathmandu practical reasons to coordinate closely with Beijing.
  • The story is part of a wider pattern in which disaster response becomes a channel for regional influence, technical data sharing and bilateral trust.
Open at SCMP Ask Claude ChatGPT

China tests a stealth drone from its new amphibious assault ship

# Policy SCMP

China's GJ-21 stealth attack drone has reportedly entered flight testing as the Type 076 amphibious ship approaches final sea trials, pointing to a carrier-like role for unmanned strike aircraft.

More
  • The GJ-21 is described as a naval version of the GJ-11 and is intended to operate from the Type 076 platform.
  • The combination suggests a shift from amphibious ships as troop-and-helicopter vessels toward distributed aviation, surveillance and strike operations.
  • The source says the drone has made few public appearances since its parade debut, so the maturity of its launch, recovery and deck-handling systems remains unclear.
  • The strategic value depends on the full system: shipborne command links, sortie generation, maintenance and resilient communications matter more than the airframe alone.
  • This is an important signal of how China is integrating unmanned systems into naval force design rather than treating drones as standalone aircraft.
Open at SCMP Ask Claude ChatGPT

Xi and Putin may revive talks on Power of Siberia 2

# Policy SCMP

The Kremlin says Xi Jinping and Vladimir Putin are expected to discuss the Power of Siberia 2 gas pipeline in Bishkek, underscoring the strategic depth of the China-Russia relationship.

More
  • The meeting would be the leaders' second face-to-face encounter of the year and comes after Putin's visit to Beijing following a US presidential visit there.
  • Power of Siberia 2 would create another major gas link from Russia toward China, but the report confirms discussion rather than a signed construction or pricing agreement.
  • For Beijing, the project could diversify pipeline supply and deepen continental energy interdependence; for Moscow, it offers an alternative market as relations with the West remain strained.
  • The timing shows how Central Asian diplomacy, energy security, and great-power signaling are being negotiated together rather than as separate commercial files.
  • The meaningful follow-up is whether the two sides resolve route, financing, capacity, and pricing questions that have kept the project politically attractive but commercially unsettled.
Open at SCMP Ask Claude ChatGPT

A Chinese industrial robot wins more than $14 million in orders

# AI 36Kr

Xuanchuang Robotics raised several tens of millions of yuan and has more than 100 million yuan, roughly $14 million, in orders for robots designed for explosive and hazardous chemical sites.

More
  • Its AEGIS architecture separates high-level task planning, motion reasoning and a millisecond-scale safety barrier, preventing an unsafe model command from reaching the actuators.
  • The company uses wheel, track and rail configurations rather than humanoid bodies because refinery and chemical-plant customers prioritize reliability, certification and access over human shape.
  • Remote operation is treated as both a safety intervention and a data-collection path toward autonomous operation; the company targets routine autonomy by 2030 and complex autonomy by 2035.
  • The source says more than 100 units have been ordered by the China National Petroleum system in Xinjiang, with initial deliveries underway at about 20 units per month.
  • The technical moat is domain data and process knowledge: valve torque, sequencing and local safety practice are rarely captured in academic datasets.
  • This is a stronger commercialization signal than a stage demo because the product is sold into a regulated, dangerous environment where payback, certification and failure containment are explicit customer requirements.
Open at 36Kr Ask Claude ChatGPT

A Shenzhen programmer's death tests the limits of Chinese workplace protection

# Culture Solidot

Authorities rejected a family's claim that a programmer who died in his office building's restroom suffered a work injury, arguing he had not reached his desk; the family is appealing.

More
  • The 39-year-old reportedly entered the building, went to a second-floor restroom, and was found unconscious more than two hours later before being pronounced dead.
  • The disputed issue is whether a workplace restroom is a reasonable extension of the job site or whether the worker must first reach the assigned workstation, a narrow rule with large consequences for families.
  • Recognition could unlock about 1.13 million yuan (roughly $159,000) plus funeral and survivor benefits, giving the legal classification direct economic weight.
  • The case exposes a recurring gap in China's labor system: formal workplace boundaries can lag behind how employees actually move through employer-controlled premises.
  • The source includes a lawyer's interpretation but no court ruling, so the agency's decision is not the final legal position.
Open at Solidot Ask Claude ChatGPT

China and Russia revisit an Arctic route through the Irtysh

# Policy SCMP

Xi Jinping and Vladimir Putin are expected to discuss a river-to-Arctic cargo route linking China through Kazakhstan and Russia, a concept that could reduce exposure to contested sea lanes.

More
  • The proposed corridor would use the 4,248-kilometer Irtysh River, which begins in China's Xinjiang region, crosses Kazakhstan, and joins Russia's Ob River before reaching the Arctic Ocean.
  • Its attraction is strategic geography: an inland waterway could complement maritime routes that are vulnerable to conflict, chokepoints, sanctions, and naval pressure.
  • The route would face major engineering, seasonal, environmental, sovereignty, and commercial constraints; the report describes talks, not an approved infrastructure project.
  • The political value may arrive before the logistics: China and Russia can use the corridor to signal deeper continental integration even if cargo economics remain uncertain.
  • The story belongs in the context of China's search for resilient trade and energy corridors across Central Asia and the Arctic.
Open at SCMP Ask Claude ChatGPT

Unitree's market debut exposes the gap between robot hype and demand

# Analysis TMTPost

Unitree's extraordinary first-day valuation surge faded sharply within days, forcing China's humanoid-robot sector to confront whether research and education sales can become industrial deployment.

More
  • The stock opened at 1,100 yuan and briefly valued the company at 444.9 billion yuan, roughly $62.7 billion, before falling to about 243.9 billion yuan, roughly $34.4 billion.
  • The listing gave investors a public benchmark for a sector previously priced mainly through private funding rounds and comparisons with future products.
  • Unitree reported revenue rising from 159 million yuan in 2023 to 1.699 billion yuan in 2025, with 2025 profit of 278 million yuan, roughly $39 million.
  • The harder fact is revenue mix: research and education represented 73.6 percent of humanoid-robot revenue in the cited 2025 period, while industry applications were 9.01 percent.
  • The company says 2025 humanoid shipments exceeded 5,500, but first-half 2026 adjusted profit fell 19.34 percent even as revenue rose 48.54 percent, suggesting scale is pressuring margins.
  • The market is now likely to demand evidence of repeat industrial orders, manufacturing yield, falling unit costs and reliable manipulation rather than demonstrations alone.
  • Unitree's valuation matters beyond one company because it is becoming the public-market stress test for China's entire embodied-intelligence investment narrative.
Open at TMTPost Ask Claude ChatGPT

Huawei pitches hybrid-cloud data infrastructure built for agents

# Policy Leiphone

Huawei says its hybrid-cloud stack now links governance, multimodal data lakes, trusted data spaces, and agent access, reflecting China's effort to make enterprise data usable without moving it all to a central cloud.

More
  • The architecture separates reliable management, efficient supply, intelligent use, and trusted circulation, connecting cloud, data-center, factory, and office data under a hybrid-cloud control plane.
  • Huawei describes cross-region disaster recovery, billion-scale metadata management, multiple data engines, and data access without relocating every source, aimed at enterprises that cannot surrender operational or industrial data.
  • The platform also uses privacy computing, APIs, sandboxes, and blockchain to create auditable data-sharing environments across organizations, a direct response to data sovereignty and provenance concerns.
  • Chinese industry examples include a power utility's multidimensional labeling system, a regional data chain covering nine provinces and 40 cities, and city-level trusted data spaces.
  • The policy and business signal is that Chinese AI deployment is being framed as a data-infrastructure problem before it is a model problem: agents need governed, current, machine-readable enterprise context.
  • The claims come from a Huawei-led conference and partner case studies, so independent measures of latency, accuracy, cost, and real production dependence are still missing.
Open at Leiphone Ask Claude ChatGPT

China spends a decade building a national map of its languages

# Culture SCMP

A new Hunan museum hall showcases dialects, minority languages, and oral culture collected across more than 1,800 sites over ten years, turning linguistic preservation into a technology project.

More
  • Visitors can speak into a recognition booth that identifies their likely hometown dialect, making a large archival effort legible through an interactive public interface.
  • The collection covers dialects, minority languages, and oral traditions, preserving material that is vulnerable to migration, schooling, media standardization, and generational change.
  • The project shows China's state capacity at its most cultural: museums, field researchers, language technology, and national data collection are being combined to document diversity while the country also promotes a common standard language.
  • The source does not specify the corpus size, licensing terms, or whether the data will be available for commercial model training, which will determine its value to speech and language AI.
  • The strategic importance is both cultural and technical: language archives can support education and preservation, but they also become infrastructure for recognition systems and future generative models.
Open at SCMP Ask Claude ChatGPT

China's vertical apps are losing users to bloated business models

# Culture TMTPost

Keep's active users and paid-content revenue continue to fall as the fitness app shifts toward merchandise, illustrating how Chinese vertical platforms can destroy their original utility while chasing broader monetization.

More
  • Keep's first-half revenue was 825 million yuan, roughly $116 million, nearly flat year on year, while online membership and paid-content revenue fell 26.9 percent.
  • Average monthly active users fell from 29.9 million in 2024 to 18.6 million in the first half of 2026; monthly subscribers fell to 2.17 million.
  • The company says low-frequency, trend-driven users left, but interviews in the article suggest committed users also object to crowded feeds, shopping features and formerly free content moving behind a paywall.
  • The same pattern may apply beyond fitness: a focused utility builds trust, then advertising, commerce and community layers make the core task harder to find.
  • The demographic mismatch is notable: the article cites older adults as more persistent exercisers while younger users are more exposed to fatigue and disengagement, yet the app is not designed around the former group.
Open at TMTPost Ask Claude ChatGPT

Bilibili turns advertising into its main growth engine

# Culture TMTPost

Bilibili's second-quarter advertising revenue rose 28% to 3.13 billion yuan, roughly $441M, surpassing value-added services for the first time as AI helps monetize its community.

More
  • The platform reported 7.94 billion yuan in quarterly revenue and 700 million yuan in adjusted profit, marking an eighth consecutive profitable quarter.
  • Automatic advertising accounted for more than 85% of ad spend, while AI-related ad revenue doubled year over year, showing AI being used primarily to improve monetization efficiency.
  • Bilibili is spending about 1 billion yuan, roughly $141M, on AI capital expenditure across video understanding, distribution, and creation rather than competing in frontier-model infrastructure.
  • The shift is culturally significant because Bilibili historically protected its community from aggressive advertising and depended more on memberships and games.
  • Aging users with more purchasing power make the platform more attractive to automotive, home, and other high-ticket advertisers, but commercial pressure can erode the user experience that created the audience.
  • Its strategy is a useful counterexample to the AI arms race: a content platform can use models to increase revenue and lower production costs without training a giant general model.
  • The open risk is that advertising grows by making the community less distinctive, while games and memberships remain too slow to offset that pressure.
Open at TMTPost Ask Claude ChatGPT

Huawei previews a more protective wing-fold tri-fold phone

# AI Leiphone

Huawei's next luxury tri-fold phone encloses its inner display when closed and uses a centered hinge layout, extending the company's lead in China's ultra-premium foldable market.

More
  • The new wing-fold form is intended to reduce abrasion and impact exposure compared with Huawei's earlier outward-facing Z-fold arrangement.
  • The company is signaling that the next competitive frontier is reliability and ergonomics: hinge precision, weight balance, crease control, and resistance to drops matter as much as screen area.
  • Huawei's prior tri-fold reportedly attracted more than six million reservations and sold nearly 500,000 units, according to the article, but those figures are company and media estimates rather than audited demand data.
  • The product's high-end materials and distinctive styling show how Chinese phone makers use mechanical novelty and luxury branding to defend premium pricing.
Open at Leiphone Ask Claude ChatGPT

Anthropic opens a hardware standard for AI-controlled laboratories

# AI New Wisdom

Anthropic's Model Hardware Standard lets Claude operate lab instruments through a common interface, while Chinese coverage emphasizes the leap from software tools to physical experimentation and the remaining safety gaps.

More
  • The supplied account describes Claude controlling robotic arms, liquid handlers, microscopes, cameras and quantum-lab laser systems through a common hardware description and API layer.
  • At QuEra, the system reportedly tested and repaired hundreds of laser faults overnight, reaching 99.3% stability versus 58% for human experts in the cited trial.
  • The architectural idea resembles a physical-world tool protocol: devices expose capabilities and constraints so an agent can compose them without custom integration for every instrument.
  • The failure mode is equally important: when a viscous protein solution foamed, the agent repeated the same action instead of recognizing a physical cause and switching wells.
  • Legacy equipment without APIs remains a major bottleneck, and continuous monitoring creates compute costs and an unresolved question of liability.
  • The project is a research preview, not an autonomous laboratory standard; safety envelopes, audit trails and human intervention rules will determine whether it scales.
Open at New Wisdom Ask Claude ChatGPT

Chinese scholars warn that AI can make Beijing sound more threatening

# Policy SCMP

A state-affiliated Chinese academic paper urges human review of AI translations because literal renderings of political language may distort Beijing's intent and worsen foreign perceptions.

More
  • The recommendation is for stronger human translation capacity and oversight as AI takes over more diplomatic and policy translation work.
  • The concern is not just wording quality: political language often relies on institutional context, calibrated ambiguity, and audience-specific conventions that a literal model may flatten into threats.
  • The story reveals an official Chinese anxiety about losing control of external signaling even while adopting AI internally, especially when machine output can harden language for foreign readers.
  • Human translators are being positioned as strategic interpreters rather than a temporary legacy layer, suggesting that high-stakes state communication will remain hybrid.
  • The missing evidence is any measured comparison of AI and human translations or a concrete policy mandate, so this is a warning about governance rather than a completed rule.
Open at SCMP Ask Claude ChatGPT

Chinese AI long-form films reach a commercial turning point

# Culture TMTPost

Three Chinese AI long films arrived within a month, moving generative video from short-form novelty into licensed, serialized storytelling and a new low-cost production model.

More
  • The first licensed AI web film reportedly exceeded an hour and was made end to end with AI, while a later franchise prequel opened with a 7.8 rating on Douban.
  • The productions retain human control over story, emotion, and performance while using AI to reconstruct younger actors, extend visual worlds, and reduce production costs.
  • China's licensing system is part of the experiment: one series is using simultaneous production, review, and broadcast rather than waiting for a fully finished season.
  • The business pitch is portfolio economics, with platforms and investors able to fund many targeted mid-budget projects instead of betting everything on a few blockbusters.
  • The technology hurdle has shifted from whether a model can render shots to whether it can preserve character identity, causal plot, and audience trust across feature length.
  • The cultural question is whether AI becomes a production instrument inside traditional studios or creates a separate creator economy with different labor and rights norms.
  • Sustained viewership, repeatable revenue sharing, and treatment of performers' likenesses will determine whether this is a durable market.
Open at TMTPost Ask Claude ChatGPT

China's family-doctor expansion makes AI follow-up a policy priority

# Policy TMTPost

A Chinese health-policy analysis argues that expanded family-doctor services and unified reimbursement are turning AI-assisted patient follow-up from an optional feature into required infrastructure.

More
  • The policy expands family-doctor work from routine chronic-disease reminders toward specialist coordination, infectious-disease education and more complex medication management.
  • The source links this to a first batch of 158 conditions receiving the same reimbursement standard across hospitals and community facilities, encouraging care to move downward into primary clinics.
  • It describes four business models: government-purchased software, deeply operated care networks, ecosystem platforms and commercially paid services.
  • One cited system lets a health manager oversee 2,600 patients with AI support, about 3.5 times traditional productivity, while another reports AI suggestions across more than 77,000 primary-care institutions.
  • The commercial question is who bears liability when AI-supported follow-up is wrong; the policy language emphasizes safety and control but the article does not define a clear allocation of responsibility.
  • The story shows how Chinese regulation can create a market by expanding mandated service scope before frontline staffing and budgets are ready.
Open at TMTPost Ask Claude ChatGPT

Tencent's Hy4 shows how model development is becoming a product loop

# AI TMTPost

Tencent compressed Hy4's release cycle while using real products, expert feedback, and infrastructure experiments as a training and evaluation loop.

More
  • The supplied analysis says Hy4 reached 64.3 on DeepSWE, up from Hy3's 28.0, and matched a leading competitor on Terminal-Bench 2.1 at 85.4.
  • The more consequential change is organizational: product users supply real prompt distributions, specialists define quality, and model engineers turn both into training data and tests.
  • Tencent says more than 50 evaluation suites combine real tasks, human grading, and product testing rather than relying on public leaderboards alone.
  • This is a classic deployment-feedback strategy, but Tencent can apply it across office, coding, gaming, finance, and security products at once.
  • Preview releases are being treated as controlled data-collection mechanisms, not merely unfinished products.
  • The approach can improve tool use and recovery from failure, but it also raises governance questions about what enterprise and consumer interactions are retained and reused.
  • The differentiator may be the feedback pipeline and distribution surface, while the base model remains in a crowded catch-up race.
Open at TMTPost Ask Claude ChatGPT

Former Chongqing airport executive placed under investigation

# Policy Jiemian

China's anti-corruption system has opened a disciplinary and supervisory investigation into a former senior executive of the Chongqing airport group.

More
  • The announcement comes from the Chongqing discipline inspection and supervision authorities and alleges serious violations, without stating specific conduct.
  • The case shows how the national inspection system reaches state-owned enterprises and their senior managers, not only government departments.
  • Such notices are intentionally sparse at the investigation stage; the supplied account gives no conclusion about guilt or the eventual penalty.
Open at Jiemian Ask Claude ChatGPT

Tencent's Hy4 leads a Chinese tech roundup on models and platform economics

# GeekPark

Chinese coverage presents Hy4 as a million-token productivity model, while the same roundup stresses that Meituan wants AI to improve its core business rather than become a token factory.

Also: Hacker News

More
  • The Hy4 figures in this roundup match the separate release account: 770 billion total parameters, 49 billion active parameters and a 1M context window.
  • The roundup reports Meituan revenue of 104.6 billion yuan, roughly $14.7 billion, up 14.4 percent, with second-quarter profit of 2.16 billion yuan, roughly $304 million.
  • It also points to a new $1.1 billion hardware-infrastructure fund from Andreessen Horowitz and rising memory costs, linking model economics to chips, storage and systems.
  • As a roundup, it combines separate events and provides less depth than the dedicated Hy4 and Meituan reports.
  • The Chinese framing treats AI as an operating layer across existing businesses, not only as a race to train the largest model.
Open at GeekPark Ask Claude ChatGPT