China Desk — 13 Aug 2026, 22:32 UTC

Zhu Rongji's political legacy: reform, centralization and social cost

# Initium 4h

Chinese commentary on Zhu Rongji's death remembers both an anti-corruption reformer and the architect of reforms that centralized fiscal power while leaving harsh memories among workers and families.

Also: SCMP, The Wire China

Zhu's legacy is unusually useful for understanding the state-market bargain that shaped modern China: fiscal and administrative centralization enabled rapid integration while concentrating the costs of restructuring.
The supplied podcast emphasizes the tension between popular memories of personal incorruptibility and criticism of the economic system he built.
Chinese coverage treats his reputation as contested history rather than a simple reformer memorial: market transformation and social dislocation remain inseparable.
The key context is that his era connected China more tightly to global markets while strengthening the central government's capacity to steer the economy.
What remains unclear from the excerpt is how official institutions will frame his legacy beyond formal mourning language.

Open at Initium Ask Claude ChatGPT

Hong Kong's national-security cases reach into bookshops and courts

# Initium 15h

Hong Kong authorities searched independent bookshops and arrested five people over alleged sedition, while a young autistic man was convicted over political graffiti and a judge retired early after repeated plagiarism findings.

The bookshop raids show how national-security enforcement now reaches cultural distribution: possession and sale of politically sensitive books can be treated as evidence of seditious intent.
The report says the 18-year-old received an 18-month probation order, the first non-custodial sentence under the city's Article 23 sedition offense, illustrating both criminalization and calibrated sentencing.
A judge's early retirement after repeated findings of judicial plagiarism adds a separate institutional signal: the judiciary is protecting formal credibility while operating inside a tightened political environment.
The cases involving independent bookstores and the earlier closures of several shops suggest a chilling effect on Hong Kong's remaining pluralist publishing infrastructure.
The account also describes a rejected judicial review by 87 mainland students over subsidized university places, with the court prioritizing residents with a close local connection.
Taken together, the month’s cases show national security, administrative allocation and institutional discipline operating across different parts of Hong Kong's governance system.
The supplied report is a monthly roundup; individual cases may still be subject to appeal or later factual clarification.

Open at Initium Ask Claude ChatGPT

DeepSeek V4 Pro's final release targets the agent bottleneck

# GeekPark 13h

DeepSeek's V4 Pro final release keeps the preview architecture but sharply improves reported software-engineering and tool-use scores, at 3 yuan per million input tokens (roughly $0.42) and 6 yuan output (roughly $0.85).

Also: Leiphone

The model remains a 1.6-trillion-parameter, 49-billion-active-parameter MoE with a one-million-token context and 384,000-token maximum output, according to the supplied report.
The claimed gains are concentrated in long-chain work: DeepSWE rises from 12.8 to 62.7, CyberGym from 52.7 to 83.3 and Terminal Bench from 72.1 to 87.9.
That pattern suggests post-training and tool-use behavior, rather than a new base architecture, are doing most of the work; the release is aimed at making agents execute rather than merely reason in text.
Hands-on testing found it could generate complete interactive programs, but a reasoning-heavy mode sometimes consumed most of the output budget thinking and truncated the actual code.
The model is not described as universally best: the supplied tests place it behind leading closed models on some knowledge and repository tasks.
The strategic advantage is price-performance. At the reported prices, DeepSeek can make long-running agent workloads economically viable even when it does not lead every benchmark.
The release also appears linked to DeepSeek Harness, suggesting the company is optimizing a model-runtime pair rather than treating the model as a standalone API.

Kingsoft brings DeepSeek V4 Pro into its office agent Leiphone

Open at GeekPark Ask Claude ChatGPT

China's FlagOS adapts a 2.4-trillion-parameter Qwen model across nine chips

# Leiphone 13h

Alibaba's Qwen3.8-2.4T-A95B MoE model was reportedly adapted on nine Chinese and foreign accelerator platforms through the open FlagOS stack, with INT8 quantization aimed at extending existing hardware life.

The model has 2.4 trillion total parameters and 95 billion active parameters, with a reported 262,144-token native context extendable to about one million tokens.
FlagOS says it completed cross-chip precision alignment and deployment validation for BF16, FP8 and INT8, including weight-and-activation quantization to reduce memory pressure.
The architectural significance is portability: a common compiler, operator library and plugin layer can turn each new model release into a multi-chip deployment event instead of a bespoke port for every accelerator.
The program reportedly reduced adaptation time to under 24 hours, which could let smaller Chinese cloud providers offer new models without waiting for a chip vendor's dedicated engineering effort.
Chinese coverage frames this as a way to preserve the value of installed accelerator fleets, especially hardware that cannot efficiently run FP8 workloads.
The performance and compatibility claims are from the project; independent throughput, power and accuracy comparisons across all nine platforms are not supplied.

Open at Leiphone Ask Claude ChatGPT

US accuses Chinese exporters of a broad transshipment scheme

# SCMP 11h

Chinese coverage foregrounds Washington's accusation that exporters routed goods through more than 40 third countries to evade tariffs, while the US framing calls it the Great Transshipment Scam and promises enforcement.

The administration's 25-page report alleges limited assembly, relabeling and rerouting have been used since 2018 to avoid US tariffs; these are allegations in the supplied account, not adjudicated findings.
The policy implication is a move from bilateral tariff enforcement toward origin tracing across third-country supply chains, raising compliance costs for manufacturers and logistics firms.
For Chinese exporters, the strategic problem is not only the tariff rate but the loss of trusted intermediary jurisdictions and the risk that minor processing will no longer establish origin.
Chinese coverage leads with the US accusation and its trade-war framing, whereas the American label presents the practice as organized fraud; neither side's excerpt supplies independent verification.
The issue matters most for companies whose global footprint depends on assembly, warehousing or invoicing outside China, especially in electronics and industrial goods.
A durable response would require clearer origin rules and evidence-based customs procedures; otherwise the dispute may accelerate supply-chain bifurcation rather than stop transshipment.

Open at SCMP Ask Claude ChatGPT

Fudan team ranks second in an AI cyberattack benchmark

# Jiemian 7h

Fudan University's Whitzard agent achieved a reported 91.2% success rate against 1,507 real vulnerabilities across 188 open-source projects, ranking second globally and first among universities.

The benchmark requires an agent to locate vulnerabilities, construct exploits and verify them inside complete codebases rather than answer isolated security questions.
The team used one base model plus a modular framework that converts program reachability into reasoning constraints, with total test calls costing under 5,000 yuan (roughly $700).
The reported result improved from 68.9% to 91.2%, with 1,374 vulnerabilities successfully handled; the framework and benchmark are said to be open source.
Chinese coverage emphasizes engineering efficiency and a domestic university outperforming several prominent corporate teams, a useful counterweight to model-brand narratives.
The main caveat is source dependence: the supplied account does not identify the benchmark's governance or provide independent replication.

Open at Jiemian Ask Claude ChatGPT

China's nuclear capacity reaches 135 gigawatts in operation and construction

# IT Home 7h

China says 57 nuclear units are operating or approved under construction, totaling 135 gigawatts; nuclear supplies nearly 5% of electricity from less than 2% of installed capacity.

The reported operating-plus-approved capacity is up from 125 gigawatts at the end of 2025, keeping China first globally in both total and construction scale.
The buildout is concentrated in coastal provinces, where nuclear already supplies more than 15% of generation in Guangdong, Fujian, Liaoning and Hainan.
Chinese official coverage presents the fleet's safety record as a strategic asset, citing no level-two-or-higher international nuclear events and a decade-leading operations index.
This is infrastructure policy as much as energy news: nuclear provides dense, predictable low-carbon power for industrial coastal economies while China expands other generation sources.
The figures are official and aggregate; they do not show project economics, construction delays or the future fuel and waste burden.

Open at IT Home Ask Claude ChatGPT

DeepSeek Harness is an open, pluggable runtime for self-evolving agents

# AI QbitAI 6h

Chinese testing finds DeepSeek Harness unusually configurable: models, tools, policies, storage and context management are pluggable, while trajectory logs expose event-level agent behavior and token costs.

Also: Leiphone, Hacker News, V2EX

The architecture is closer to an agent runtime than a polished assistant: developers can swap components, define presets and inspect raw event traces rather than only a summarized chat.
The trajectory view addresses a real production problem: long tool chains become opaque, so engineers need replayable state and cost visibility to debug failures.
The review reports more than 100 built-in plugins and high cache-hit rates in testing, but also notes missing desktop packaging and weaker interface polish than established coding agents.
The author's central judgment is that Harness is designed for developers who want to modify the execution loop, not for users seeking a finished application.
Chinese coverage frames the release as a platform for community experimentation and self-evolution, while the main uncertainty is whether that openness can become dependable enterprise infrastructure.
Its significance rises if the runtime becomes the reference environment for DeepSeek's post-training and third-party agents; the supplied review does not establish that yet.

Open at QbitAI Ask Claude ChatGPT

Lenovo's AI server business drives a record quarter

# Signal QbitAI 15h

Lenovo reported record first-quarter revenue of 183.4 billion yuan (roughly $25.8 billion) and adjusted profit above $1 billion, with infrastructure revenue nearly doubling year on year.

Infrastructure Solutions revenue reached 57.9 billion yuan (roughly $8.2 billion) with a 9.1% operating margin, making the server business a new profit engine rather than a low-margin adjunct.
Lenovo says AI-related revenue rose 60% to 63.4 billion yuan (roughly $8.9 billion), about 35% of group revenue, while AI services grew 141%.
The strategic model is hybrid: global sourcing and local delivery let Lenovo sell AI infrastructure while absorbing supply-chain and component-price volatility across regions.
The company also reports that its PC agent has reached a 40% quarterly weekly-active rate and that Qira conversion is about 28%, but those metrics are company-defined and not independently audited.
The results show Chinese technology manufacturers capturing value from the infrastructure wave even when they do not control the leading accelerator designs.
The key question is whether the infrastructure margin can persist once hyperscalers normalize spending and memory and accelerator supply loosen.

Open at QbitAI Ask Claude ChatGPT

What DeepSeek's unusual business and research thesis reveals

# Analysis ChinaTalk 12h

ChinaTalk's reading of leaked investor minutes portrays DeepSeek founder Liang Wenfeng as pursuing automated learning and general intelligence as a long-term mission, with open models serving a business strategy beyond licensing.

The essay describes a founder who treats hardware limits as a constraint on implementation rather than a reason to narrow the research horizon, a stance that differs from most commercially timed Chinese AI companies.
Its business question is central: open weights can still generate revenue through enterprise services, integration and demand for capable inference, rather than by selling exclusive model access.
The political-economy angle is that DeepSeek's Hangzhou base combines private capital, unusually patient management and a national hardware constraint with a global open-source posture.
The supplied account is analysis based partly on leaked minutes, so interpretations of Liang's motives should not be treated as confirmed company policy.
The comparison with Western lab founders is useful only if it preserves the institutional difference: DeepSeek's choices are shaped by Chinese capital, compute restrictions and an ecosystem that values deployable open models.
Whether this thesis produces durable advantage depends on sustained post-training progress and a business model that can fund frontier research without closing the weights.

Open at ChinaTalk Ask Claude ChatGPT

China's Coolfly unveils a standing personal eVTOL

# AI QbitAI 12h

Chinese company Coolfly unveiled standing and seated single-person electric aircraft, using a dual-redundant flight controller with six IMUs per independent control unit.

The standing Urban is designed around a human-machine system whose center of gravity changes as the pilot moves, requiring the controller to compensate for body motion and disturbances in real time.
Coolfly's NA80 architecture uses two independent flight-control units, cross-checks multiple inertial sensors and reportedly supports millisecond-level fault detection and handover.
The seated Dream Pro is reported to carry up to 120 kilograms at 100 kilometers per hour for as long as 40 minutes with a 70-kilogram pilot; the standing model's stated endurance is about 11 minutes.
The company says it has nearly 100 paid overseas deposits and insurance arrangements, but deposits and regulatory eligibility are not evidence of safe commercial operation.
The Chinese low-altitude-economy push is moving from drones toward passenger products, where certification, training, insurance and public-risk management will determine whether the category scales.

Open at QbitAI Ask Claude ChatGPT

Chinese missile AI reportedly tracks stealth-fighter heat signatures

# Analysis SCMP 8h

Chinese researchers reportedly trained an AI system to distinguish F-22- and F-35-like infrared signatures from decoy flares with more than 90% accuracy, pointing to an electronic-warfare use case.

The technical problem is classification under deception: engine exhaust and skin-heating signatures differ from flares, so an interceptor can use temporal and spectral patterns rather than simply chase the brightest infrared source.
The reported result concerns target recognition for heat-seeking missiles, where false locks and decoy rejection matter as much as raw detection range.
This is a Chinese military research story, and the strategic implication is that stealth reduces radar visibility without eliminating thermal observables.
The supplied excerpt does not identify the model architecture, dataset size, test conditions or whether the result was demonstrated outside simulation.
The result matters more if it survives sensor noise, changing aspect angles and adversarial countermeasures; the stated accuracy alone does not establish operational readiness.

Open at SCMP Ask Claude ChatGPT

Pony.ai moves its fourth-generation autonomous truck toward production

# Analysis Leiphone 18h

Chinese self-driving company Pony.ai unveiled a fourth-generation heavy truck and targets operating 500 to 1,000 vehicles within two to three years, shifting from demonstrations toward fleet economics.

The company says the new window opened because vehicle integration, safety redundancy, operating costs and deployment scenarios are mature enough to justify scale, not because the underlying autonomy problem is solved.
Its stack reuses more than 80% of code and over 90% of hardware from other products, according to company estimates, while adapting to charging, tolls, service areas and mixed traffic on long routes.
Pony.ai is prioritizing long-haul freight, western China bulk-commodity routes and ports, where driver shortages and round-the-clock operations create a clearer economic case than ordinary passenger transport.
The two reported business models are technology-as-a-service and transport-as-a-service, with the company eventually hoping to retreat to a software and vehicle-systems role.
This is a useful Chinese commercialization signal: autonomous trucking is being evaluated in cost-per-ton-kilometer terms, not just miles driven without intervention.
The target is ambitious and company-reported; fleet uptime, safety incidents, regulatory permissions and customer economics will determine whether production becomes a business.

Open at Leiphone Ask Claude ChatGPT

Didi's China mobility orders rise for a 14th straight quarter

# Signal Leiphone 12h

Didi's core platform handled 5.052 billion second-quarter orders, up 13.2%, while China mobility grew 8.1% and total transaction value reached 133.9 billion yuan (roughly $18.9 billion).

The results show a mature Chinese platform still expanding through service segmentation rather than only subsidized volume: premium vehicle categories and standardized service are emphasized.
Didi says its China mobility business generated 90.4 billion yuan (roughly $12.7 billion) in quarterly transaction value, while its international business grew faster and remained profitable.
The company reports that its AI assistant uses more than 90 service tags and a multi-agent architecture to turn preferences into matching decisions, improving satisfaction in core scenarios by nearly 10 percentage points.
The platform is broadening from rides into delivery, travel, charging and vehicle-owner services, illustrating how Chinese internet firms use a dominant entry point to build local-life ecosystems.
These are company-reported figures, and the report does not show take rate, driver churn or whether the service upgrades improve Didi's own margins.

Open at Leiphone Ask Claude ChatGPT

Black-box attacks expose weaknesses in robot navigation systems

# AI Leiphone 13h

A Hong Kong university team developed AdvNav, a black-box attack that perturbs visual input and reportedly causes vision-language navigation agents to fail without access to model weights or gradients.

AdvNav uses trajectory-level and action-level feedback, adaptive perturbation strength and genetic search to optimize attacks under limited query budgets.
On the R2R benchmark, the reported attack success rates against Transformer-based HAMT and language-model-based MapGPT reached 49.70%, 65.96% and 87.30% across tested settings.
The architectural lesson is that stronger reasoning can increase dependence on perception: a small visual error can propagate through a long sequence of decisions before the agent has a chance to recover.
For deployed robots, the threat is not merely adversarial images in a lab; camera contamination, occlusion and environmental changes can create similar distribution shifts.
The supplied study demonstrates vulnerability on benchmark systems, not operational compromise of a commercial robot, and the excerpt does not establish physical-world transfer.

Open at Leiphone Ask Claude ChatGPT

Tencent's AI spending surge is also a bet on scarce memory

# Signal 36Kr 13h

Tencent's second-quarter revenue reached 204.8 billion yuan (roughly $28.8 billion), while capital spending jumped to 52.8 billion yuan (roughly $7.4 billion), turning free cash flow negative as it secures AI infrastructure.

The analysis argues that Tencent is buying not only accelerators but high-end memory and storage, because inference throughput depends on moving model state as well as multiplying matrices.
The company reported research spending of 27.3 billion yuan (roughly $3.8 billion), and the article links the spending to larger future models, internal AI products and rising agent workloads.
This is a useful Chinese corporate signal: a cash-rich platform is willing to absorb short-term financial pressure to reserve scarce hardware rather than wait for AI demand to become visible in revenue.
The storage thesis is partly the author's interpretation; the supplied text does not disclose Tencent's exact memory inventory, supplier contracts or product-level AI returns.
The strategic risk is capital intensity: if model efficiency improves faster than usage, Tencent may have overbought infrastructure, but if demand keeps rising, delayed memory capacity could become the binding constraint.

Open at 36Kr Ask Claude ChatGPT

China's STEM students turn to Singapore as Western options narrow

# Culture SCMP 32m

As US-China tech rivalry and tighter immigration rules reshape study routes, Chinese science and engineering students are increasingly looking to Singapore instead of the traditional Western destinations.

The shift reflects a strategic change in talent mobility: access to laboratories, visas and technology ecosystems is becoming part of the geopolitical competition over skilled people.
Singapore offers proximity to China and strong links to Western research, but it cannot fully replace the scale and depth of the four traditional destinations.
Chinese coverage frames this as a fracture in a once-reliable path for the country's top technical students, not merely a change in university preferences.

Open at SCMP Ask Claude ChatGPT

Chinese EV exports increasingly rely on local operating ecosystems

# Analysis Leiphone 12h

In Brazil, Didi's 99 and rental company Movida plan discounted EV access for at least 50,000 drivers, showing Chinese car exports being connected to local finance, charging and ride-hailing networks.

The model addresses the adoption bottleneck after a vehicle reaches a foreign market: drivers need access to capital, insurance, charging and enough utilization to make the economics work.
The partners expect the rental program to reduce vehicle costs by about 25%; the article says the wider sustainable-mobility alliance has attracted 4.1 billion Brazilian reais, or about 5.36 billion yuan (roughly $755 million), in investment.
The strategic shift is from exporting finished cars to exporting an operating system around them, with Chinese firms using platforms and supply-chain partners to create demand.
Didi's Brazilian network is described as spanning rides, delivery, payments and banking, giving Chinese manufacturers a local distribution and usage channel rather than a one-off showroom presence.
The claims come largely from company and industry sources; the missing evidence is fleet utilization, driver income after rental costs and long-term vehicle residual values.

Open at Leiphone Ask Claude ChatGPT

Satellite images suggest a 50% larger missile load for China's Type 054B

# Analysis SCMP 9h

Satellite imagery suggests a new Type 054B frigate under construction in Shanghai may have a larger vertical-launch system than the first two ships, increasing missile reserve and air-defence capacity.

The evidence comes from imagery dated June 28 and a comparison of the bow launch-system shape with two already commissioned vessels.
A larger vertical-launch system would improve magazine depth and potentially permit a broader mix of surface-to-air and strike weapons, though the exact cell count is not provided in the excerpt.
The change suggests Chinese naval design is iterating after initial production rather than freezing the class's configuration.
This is an open-source intelligence inference, not an official specification; imagery cannot by itself establish the installed weapon load or readiness.
The strategic significance lies in cumulative fleet capacity: modest per-ship changes matter when a class is built at scale.

Open at SCMP Ask Claude ChatGPT

Robot bodies may be the overlooked bottleneck in embodied AI

# Analysis Leiphone 17h

A Chinese robotics analysis argues that reliable, open robot bodies matter as much as models and data, because latency, drift, interfaces and synchronized sensors determine whether algorithms work on real machines.

The article describes a useful systems boundary: a model can only express its capability through actuators, control loops, sensing and a body that can operate continuously without drifting.
It highlights a shared control interface for teleoperation and model inference, synchronized camera and state logging, and direct deployment of training data back onto the same robot as practical ecosystem advantages.
A competition result is used as evidence that the TA2 body can sort 213 items, assemble 121 parts and move 29 boxes in 15-minute trials, though the performance claims come from the company and event context.
The strategic implication is that open, dependable robot platforms can become public infrastructure for model teams, reducing months of driver and kinematics integration work.
Chinese coverage is pushing back against the fashionable assumption that the remaining bottleneck is only intelligence or data; physical reliability may be the shortest pole.

Open at Leiphone Ask Claude ChatGPT

A Chinese startup designs a knee exoskeleton for downhill hiking

# Analysis Leiphone 19h

WudaX's Hik 1 targets downhill knee fatigue rather than the hip-assist market's usual uphill pitch, using force-aware control and potential regenerative braking to support both directions.

The engineering choice follows the load path: knee assistance can support extension uphill and damp impact downhill, whereas hip systems mainly help lift the leg and offer little downhill braking.
The company says downhill negative work could recharge the battery, though the supplied article presents this as a theoretical advantage rather than measured field performance.
Its control approach reportedly follows the user's exertion directly instead of first classifying gait into preset modes, aiming to reduce lag and the feeling of fighting the machine.
The product illustrates a Chinese consumer-robotics market still exploring fundamental architectures, not merely competing on motor count, weight and battery capacity.
The remaining questions are comfort, failure behavior on uneven terrain, power recovery in real use and whether the less obvious assistance is easy to sell to consumers.

Open at Leiphone Ask Claude ChatGPT

What the Mecca defence pact means for China

# Analysis SCMP 1h

Chinese analysis presents the Saudi-Turkey-Pakistan pact as a possible shift in Middle East security, while the accompanying expert view stresses that it remains far from a NATO-style alliance.

The agreement reportedly treats an armed attack on one signatory as an attack on all three, linking Saudi capital, Turkish military reach and Pakistan's nuclear capability.
The strategic question for Beijing is whether a more autonomous regional security network reduces US influence or instead creates another unstable alignment China must navigate.
The pact's practical force remains unclear because the members have divergent interests and limited shared military capacity.
Chinese coverage stresses the architecture-changing ambition; the more skeptical account emphasizes the gap between political symbolism and credible mutual defence.
It matters more if the mechanism develops joint planning and deployable capabilities, rather than remaining a declaration of deterrence.

Experts say the Mecca pact is far from an Islamic NATO SCMP

Open at SCMP Ask Claude ChatGPT

China's maker-tool brands chase global growth but hit practical limits

# Analysis Leiphone 14h

Chinese maker-tool companies are pushing 3D printers, laser engravers, desktop CNC and UV printers overseas, but price, safety, consumables and ease of use still constrain mass adoption.

The article estimates the global DIY maker-tool market at $18.9 billion in 2025, rising to $36.2 billion by 2033, but these are third-party market estimates cited by an industry event.
Chinese brands have made 3D printing more consumer-friendly through faster, largely turnkey machines; the next competitive frontier is multicolor printing and adjacent tools.
Desktop CNC has a different systems problem from 3D printing: dust, noise, fixturing, tool wear and CAM workflows make the product much harder to use in a home.
UV printers face a reliability bottleneck in ink and nozzle compatibility, showing how an apparently attractive consumer product can fail at consumables rather than core imaging.
The strategic question is whether Chinese manufacturers can lower imported component costs and turn specialist machines into ordinary household tools without creating unacceptable safety and maintenance burdens.

Open at Leiphone Ask Claude ChatGPT

Insta360 tries to reinvent the panoramic camera after a decade

# Analysis Leiphone 17h

Chinese camera maker Insta360's X6 combines panoramic, gimbal and first-person tracking modes with AI editing and 3D time-capsule features as hardware improvements approach practical limits.

The X6 reportedly uses a custom 1.1-inch square sensor, 8K video at 50 frames per second and a three-chip imaging architecture, but the more important change is integration of several camera forms into one workflow.
AI Director targets the central panoramic-camera problem: users capture everything and must later find the useful perspective, so editing and reframing become part of the product rather than a separate desktop task.
The company's history shows a familiar Chinese hardware playbook: discover an unexpected user behavior, build a specialized product around it, then expand from a niche enthusiast market into everyday use.
The strategic challenge is now category creation again. Better sensors alone will not sustain growth if users still think in terms of ordinary cameras and phone video.
The supplied article offers a strong product narrative but no independent shipment, margin or global market-share figures for the new model.

Open at Leiphone Ask Claude ChatGPT

Chinese startup raises funding to build physical-AI data infrastructure

# AI QbitAI 10h

Chinese embodied-AI data company Scaleforce raised several tens of millions of yuan (roughly $1-7 million) to develop its MatrixOS platform and real-world data-production network.

Physical agents need synchronized visual, tactile, joint-trajectory and environmental data that cannot be obtained by scraping the web, making collection infrastructure a distinct bottleneck from model training.
The company says MatrixOS spans collection, processing and deployment, and that it has begun a large customer order and university collaborations; these are company-reported claims.
The wider Chinese funding boom shows capital moving from robot bodies toward the data and operating layers needed to make them useful.
The supplied figures about a 99% data shortfall are industry estimates, not an independently established measurement.
The real test is whether the platform can produce repeatable, task-relevant data cheaper than robot makers can build their own collection operations.

Open at QbitAI Ask Claude ChatGPT

Chinese robotics supplier raises nearly $14 million to scale micro screws

# AI 36Kr 21h

Chinese precision supplier NOUSBOT raised nearly 100 million yuan (roughly $14 million) to expand production of miniature planetary roller screws used in robot hands and linear actuators.

The company reports a 1.5 mm screw diameter, 5.5 mm nut diameter, C5 accuracy, plus or minus 0.01 mm repeatability and 50 newtons of load capacity.
Planetary roller screws provide self-locking force and can reduce motor heating and continuous power draw compared with some rotary-joint approaches, making them attractive for compact robot joints.
NOUSBOT says its Suzhou plant is already at full capacity with annual output of three million units and aims for 10 million after the financing; these are company targets, not verified orders.
The broader signal is that humanoid-robot competition is pulling demand into obscure precision components, including wrist, neck and facial actuators rather than only hands and hips.
Reliability is the real barrier: the company says it tests toward one million to three million cycles, but the supplied account does not provide independent lifetime qualification.

Open at 36Kr Ask Claude ChatGPT

Chinese travelers mock error-filled English signs in public spaces

# Culture SCMP 14h

A mistranslated Gansu highway sign, rendered as open water rooms, triggered another Chinese online debate about poor English and spelling in public facilities.

The issue is small but revealing: public-facing translation is treated online as a proxy for administrative competence and China's ability to present itself to foreign visitors.
The report says similar errors have appeared at airports and other facilities, suggesting the problem is systemic procurement and review rather than one typo.
The story also shows how social media turns a local embarrassment into a national discussion about internationalization and service quality.

Open at SCMP Ask Claude ChatGPT

Chinese AI research is moving from scale toward reasoning and control

# Analysis Leiphone 13h

A Chinese preview of IJCAI 2026 argues that Chinese companies are emphasizing knowledge representation, planning, optimization, explainability and safety as brute-force scaling delivers diminishing returns.

The article says the conference accepted 990 papers from more than 6,500 submissions and that Chinese institutions remained prominent despite the domestic rating downgrade; these are reported conference figures.
It identifies 148 accepted papers across reasoning, planning, constraint solving and search, framing symbolic methods as roughly one-fifth of the main track.
The industry angle is practical: fraud detection, medical decisions, logistics and industrial control need predictable behavior and explanations, not only fluent outputs.
Chinese coverage presents the conference as evidence of a shift from parameter competition toward technical control and deployment in high-risk settings.
The supplied article is partly a pre-event promotional preview, so its claims about company research presence and future sessions require verification at the conference.

Open at Leiphone Ask Claude ChatGPT

Chinese small exporters move into AI and robotics supply chains

# Signal 36Kr 19h

A survey of Chinese small exporters suggests they are finding overseas opportunities in robot frames, AI-server cooling parts and medical packaging rather than competing only in finished products.

The pattern shows industrial upgrading reaching smaller firms through component supply, not just headline manufacturers of robots and models.
Cross-border payment firms are responding with local collection, currency conversion and automated compliance; XTransfer says 72 AI agents now assist onboarding, transaction verification and anti-money-laundering work.
The article reports more than one million customer companies and local collection coverage in nearly 60 countries and regions, but these are company figures rather than audited market data.
The strategic implication is that export controls and geopolitical fragmentation do not simply reduce China's exports; they also push firms toward new product categories and non-Western markets.
The missing evidence is the share of revenue or export value actually coming from AI and robotics components.

Open at 36Kr Ask Claude ChatGPT

Chinese game platforms shift from making games to finding players

# Culture 36Kr 19h

At ChinaJoy, Alibaba Cloud and TapTap said generative tools have made game prototyping easy enough that the next bottlenecks are distribution, moderation and quality control.

TapTap says users created more than 4,000 games in the first half of the year, but many cannot find players because Chinese platform traffic is already heavily priced and concentrated.
The infrastructure challenge has moved from model access to agent execution: sandboxes, tool calls, identity, auditing and content evaluation determine whether generated games can safely reach users.
This is an important consumer and labor signal. AI removes some programming barriers, but it does not remove the platform's gatekeeping role or create demand for an unlimited supply of mediocre content.
The speakers' warning that useful tools must not be immediately swallowed by foundation models points to a durable product boundary around distribution, community and governance rather than code generation.

Open at 36Kr Ask Claude ChatGPT

iFlytek pitches a full enterprise-agent stack across seven workflows

# AI QbitAI 12h

Chinese AI vendor iFlytek launched enterprise agents for management, IT, supply chain, sales, manufacturing, analysis and overseas operations, shifting its pitch from isolated tools to workflow delivery.

The announced stack includes agentic retrieval, document review, procurement monitoring, meeting analysis and an agent appliance, linking models to enterprise data and existing systems.
The company reports speech recognition across more than 100 languages, synthesis across more than 120 and service nodes in Singapore, Mexico, Ireland, Brazil and Riyadh.
Chinese enterprise AI vendors are converging on the same strategic claim: value comes from completing auditable business processes, not from winning a chat benchmark.
This is largely a product announcement, and the supplied text offers limited independent evidence of deployment, revenue or customer retention.

Open at QbitAI Ask Claude ChatGPT

A Chinese tech roundup mixes DeepSeek, Unitree and TikTok developments

# Leiphone 22h

A Chinese roundup leads with DeepSeek V4 Pro, Unitree's oversubscribed IPO and a US decision on TikTok, illustrating how AI models, robotics finance and platform politics are being consumed as one domestic tech cycle.

Also: 36Kr

The DeepSeek section repeats the reported V4 Pro release, with one-million-token context and pricing of 3 yuan per million input tokens (roughly $0.42) and 6 yuan output (roughly $0.85).
The Unitree section reports a 0.0181% online lottery success rate and more than 9.78 million applicants, a sharp signal of Chinese retail enthusiasm for robotics rather than evidence of future profitability.
The TikTok item says the US government restored use on federal devices after a restructuring that gave Oracle and Silver Lake roles while ByteDance retained a minority stake; the supplied account presents this as a regulatory-security judgment.
As a roundup, the item is useful mainly for Chinese editorial prioritization: domestic readers see model releases, robot investing and platform geopolitics as connected indicators of national technology momentum.
Several claims are attributed to other outlets and are not independently established in the supplied text.

China's daily tech roundup pairs DeepSeek with a new AI startup boom 36Kr

Open at Leiphone Ask Claude ChatGPT

Dreame's wet-dry vacuum reaches first place in Australia

# Signal Leiphone 13h

Dreame says its wet-dry vacuum held more than 25% of Australia's market in June, adapting features such as long runtime, pet-hair handling and heated self-cleaning to local homes.

The company reports a mature localization playbook: it maps product design to large homes, mixed flooring and high pet ownership, then demonstrates the features through major retailers.
Its claimed 25% share comes from GfK monitoring, but the supplied text does not identify the category definition or competitors included.
The story is a useful example of Chinese hardware brands competing abroad through channel execution and local use-case design, though it is not a major strategic event.

Open at Leiphone Ask Claude ChatGPT

Xiaomi details stricter safety targets for its Dragon Armor battery

# Signal IT Home 7h

Xiaomi says its Dragon Armor battery exceeds national requirements for impact, thermal runaway, electrical isolation, water resistance and traceability, though the claims come from the company's own Q&A.

The stated bottom-impact test uses a 25 mm striker at 500 joules, compared with a cited national benchmark of 30 mm at 150 joules; Xiaomi also requires the pack to remain functional after impact.
The company says a single-cell thermal runaway at 55 degrees Celsius and high state of charge must not propagate, and that a mechanical high-voltage disconnect acts within milliseconds.
The report describes IPX8 and IPX9K protection and cloud monitoring, but those claims are not independent validation of real-world crash or flood performance.
The broader signal is that Chinese automakers are competing through quantified safety envelopes and data-retention promises, not only range and charging speed.
The missing information is third-party test evidence and a clear distinction between the battery package and the specific vehicle trims mentioned.

Open at IT Home Ask Claude ChatGPT

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	0/20
QbitAI	200	10/10
36Kr	200	5/30
Solidot	200	10/20
Leiphone	200	19/20
GeekPark	200	2/30
IT Home	200	12/60	+42 over cap
SCMP	200	10/50	+9 over cap
Initium	200	5/15
Jiemian	200	8/30	+10 over cap
China Digital Times	200	0/15
The Wire China	200	1/50
Techmeme	200	2/15
Hacker News	200	1/20
V2EX	200	8/50	+39 over cap

DeepSeek spent two years being the floor under everyone's prices. On Monday it stops.

What Beijing is signalling

The AI race

The vibe

Read differently

Threads we are pulling