China Desk — 29 Aug 2026, 22:36 UTC

Why local model deployments drift: attention kernels and KV precision

# QbitAI 9h

A long-context experiment found that changing attention kernels or compressing KV caches can flip greedy tokens and break tool calls even with identical weights, exposing a deployment risk often hidden by short benchmarks.

The experiment captured more than 100,000 tokens from a Qwen 3.6 27B workflow on an RTX Pro 6000 and compared full-vocabulary logits rather than only final answers.
FlashAttention 2, Flash Inference and Triton Attention produced identical early outputs but diverged as context length grew, pointing to floating-point accumulation differences in prefill kernels.
With BF16 weights and Triton attention, BF16 KV cache remained stable, INT8 showed recoverable flips and INT4 produced sharply rising long-context divergence.
A concrete failure changed a Cisco interface name inside a tool call, after which later commands followed the wrong path; the experiment held token history fixed to prevent error propagation during comparison.
The weight-quantization tests found a community INT8 W8A16 build more top-1 consistent than the tested official FP8 and NVFP4 variants, partly because activations remained BF16 and sensitive layers were excluded.
At roughly 88,000 tokens, the tested NVFP4 path approached a 50 percent top-1 flip rate, while tensor parallelism itself produced non-monotonic behavior: one, two and four GPUs did not fail in the same way.
The practical implication is architectural: inference correctness depends on the entire kernel, cache, quantization and parallelism stack, not merely on checkpoint identity or advertised precision.

Open at QbitAI Ask Claude ChatGPT

Tencent open-sources Hy4 with a million-token context

# GeekPark 16h

Chinese coverage presents Tencent's Hy4 preview as a production model for long software and office workflows, while US forum reaction is more skeptical and focused on whether capability changes practical economics.

Also: Hacker News

Hy4 expands from 295 billion to 770 billion total parameters, activates 49 billion and extends context from 256K to 1M tokens, according to Tencent.
Tencent prices it at 6 yuan per million input tokens, roughly $0.85, and 18 yuan per million output tokens, roughly $2.54, with cached input at 0.3 yuan, roughly $0.04.
The model is integrated into WorkBuddy, CodeBuddy, Yuanbao and ima, and is available through Tencent Cloud and OpenRouter, making distribution part of the release.
Tencent says 163 experts blind-tested 203 engineering tasks and gave Hy4 an average 2.99 out of 4, slightly above the cited Kimi and GLM scores.
The company also says Hy4 participated in optimizing its own training methods, data strategy, evaluation and low-level operators, raising inference throughput 31.8 percent versus a baseline.
Tencent's demonstrations include cross-document expense auditing and browser-based game creation, but one test misclassified the effective date of a policy version, showing why human review remains necessary.
The strategic move is to turn a large open model into a low-cost work engine inside Tencent's product ecosystem, competing on workflow reliability and distribution rather than parameter count alone.

Open at GeekPark Ask Claude ChatGPT

A Shenzhen worker's death exposes the limits of China's injury rules

# Solidot 12h

A Shenzhen programmer who died in an office-building restroom was denied workplace-injury status because he had not reached his desk, prompting a family appeal and a dispute over how work is defined.

The source says the 39-year-old entered the building, asked security for the restroom location and was later found unconscious; doctors declared him dead after emergency treatment.
The labor authority reportedly relied on the fact that he had clocked in but had not gone to his workstation or performed work-related activity.
The family is appealing because the restroom is part of the workplace's ordinary physical environment, while the ruling treats the path to work as a narrower legal category.
The reported potential compensation exceeds 1.13 million yuan, roughly $159,000, plus funeral and survivor benefits, giving the classification a direct economic consequence.
The case illustrates a structural tension in Chinese labor governance: formal attendance systems can be easier to administer than the messy boundaries of a modern workplace.
Its broader relevance is to employers building attendance, safety and incident systems; digital records may document presence without resolving whether an employee was legally at work.

Open at Solidot Ask Claude ChatGPT

Could the Irtysh become China's Arctic shipping gateway?

# SCMP 36m

Chinese coverage presents a China-Russia river-to-Arctic route as a hedge against vulnerable sea lanes, but its huge engineering and diplomatic dependencies make it far from a practical corridor.

The 4,248-kilometer Irtysh begins in Xinjiang, crosses Kazakhstan and joins Russia's Ob River before reaching the Arctic Ocean.
The strategic appeal is obvious: an inland route could reduce dependence on maritime chokepoints exposed to conflict or blockade.
The project would require coordinated dredging, locks, ports, ice management and customs arrangements across three countries; the source does not establish that such work has been approved.
China's interest is not simply transport capacity but optionality in a world where shipping insurance, naval risk and sanctions can abruptly reshape trade routes.
The route matters more if climate change extends the navigable season and if Russia and Kazakhstan align on infrastructure; otherwise it remains a geopolitical concept rather than a near-term supply-chain substitute.

Open at SCMP Ask Claude ChatGPT

Chinese scholar warns AI translation can make Beijing sound harsher

# SCMP 16h

Chinese academic commentary urges human review of AI translations of political language, acknowledging that literal machine output can distort how Beijing's intentions are perceived abroad.

The state-affiliated article calls for stronger translation capacity and safeguards against AI-generated distortions in policy messaging.
The concern is not ordinary fluency but strategic meaning: a technically accurate rendering can still sound more threatening when it carries different assumptions about agency, certainty or coercion.
China's response is institution-building rather than abandoning automation, with human translators positioned as the final interpretive layer.
The story reflects a broader government problem with generative systems: they can scale official communication while also amplifying ambiguity, mistranslation and reputational risk.
For foreign companies and policymakers, the implication is that machine-translated Chinese policy text should not be treated as a neutral representation of intent.
The issue also reveals a tension in China's AI governance: automation is encouraged for efficiency, but politically sensitive outputs remain subject to centralized human control.

Open at SCMP Ask Claude ChatGPT

MiniMax grows fast while discounts and compute deepen its losses

# TMTPost 14h

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.

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 14h

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.

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 15h

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.

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

China reports 2.45 million PFLOPS of installed AI capacity

# IT Home 8h

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

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.

China's national AI capacity reaches 2.45 million PFLOPS Jiemian

Open at IT Home Ask Claude ChatGPT

A returning researcher becomes a Tsinghua AI class adviser

# AI QbitAI 9h

Tsinghua has appointed returning researcher Xu Mengdi as an adviser for its elite AI cohort, illustrating how Chinese universities are recruiting globally trained robotics researchers into undergraduate training.

The 2026 cohort is reported at about 90 students, including roughly 27 top informatics competitors and other olympiad-level entrants.
Xu studied vehicle engineering at Tsinghua before moving through robotics and machine learning programs in the United States and returning from Stanford in 2025.
Her research centers on adaptable generalist robots, arguing that scaling models and datasets cannot cover the near-infinite state space of the physical world.
Her RoboTool work lets robots infer tool use and executable plans from task goals and scene constraints rather than fixed action scripts.
The appointment connects elite undergraduate education to embodied AI research, with adaptation from demonstrations and reusable task structure as the conceptual bridge.

Open at QbitAI Ask Claude ChatGPT

OpenClaw's fad fades as Chinese vendors absorb its agent model

# Culture QbitAI 10h

Chinese coverage argues that OpenClaw has not died but has been commoditized: its always-on local-agent ideas now appear inside products from Tencent, Baidu, NetEase, Zhipu and ByteDance.

The article says OpenClaw grew beyond 380,000 GitHub stars by August despite social attention moving elsewhere, suggesting usage and visibility have diverged.
Its earlier appeal was a persistent agent with access to email, browsers, terminals, calendars and chat tools, often controlled remotely.
Chinese vendors reportedly launched more than 30 derivative products, shifting the value from the original brand to one-click deployment, bundled skills and domestic ecosystems.
The article's central distinction is between an agent and its harness: context management, permissions, sandboxes, long-running jobs and tool orchestration are becoming packaged platform capabilities.
This is a useful China-market pattern: open-source prototypes can create a consumer category that large platforms quickly absorb, leaving the originator with influence but less direct attention.

Open at QbitAI Ask Claude ChatGPT

Chinese war game puts the chance of coastguard weapons use at 0.3%

# Policy SCMP 8h

A Chinese maritime simulation reportedly found an extremely low weapons-use probability in a tense coastguard encounter, highlighting Beijing's preference for controlled escalation in disputed waters.

The scenario has a coastguard ship shielding fishing boats while fast small craft probe around it, forcing operators to distinguish harassment from an attack.
The reported 0.3 percent figure comes from a war-game analysis and should not be read as a forecast of actual incidents.
The strategic value lies in the decision architecture: Chinese researchers are modeling restraint, ambiguity and escalation thresholds rather than only combat performance.
Such work fits China's broader use of coastguards and civilian maritime actors to press claims below the threshold of open naval conflict.
The number would matter more if the researchers disclose assumptions, repeated trials and how weapons-use decisions were scored; the supplied account does not provide those details.

Open at SCMP Ask Claude ChatGPT

China sends nine tunnel rescuers to Nepal flood zone

# Policy SCMP 17h

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

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's decade-long project maps dialects and minority languages

# Culture SCMP 13h

A Chinese language-resource project has surveyed more than 1,800 sites over ten years, combining preservation with speech recognition that can identify a speaker's regional origin.

The collection covers dialects, minority languages and oral culture, and is being displayed through an interactive voice-recognition installation in Hunan.
The project treats language as both cultural heritage and machine-readable data, creating a corpus that can support education, research and speech technology.
The scale matters because China's linguistic diversity is difficult to capture through standard Mandarin-focused datasets.
The story also reveals a characteristic state role: cultural preservation, national data collection and technology demonstration are being pursued through one infrastructure project.
A machine that identifies region from a few sentences is impressive, but the article does not establish accuracy across accents, bilingual speakers or underrepresented languages.

Open at SCMP Ask Claude ChatGPT

Xi and Putin may revive talks on Power of Siberia 2

# Policy SCMP 9h

The Kremlin says Xi Jinping and Vladimir Putin are expected to discuss the Power of Siberia 2 gas pipeline, making the project a test of China-Russia energy alignment and bargaining power.

The leaders are expected to meet in Bishkek, with the pipeline issue discussed bilaterally, according to a Russian presidential aide cited by the source.
The project matters because it would deepen China's access to Russian gas while giving Moscow an outlet less exposed to European demand and sanctions.
A meeting is not an agreement: the source does not establish a price, construction timetable or final investment decision.
China's leverage is strongest while Russia needs eastern export routes; Russia's leverage rises if alternative Asian buyers or infrastructure constraints narrow Beijing's options.
The story belongs to the broader pattern of practical alignment without a formal alliance: energy, trade and diplomatic coordination can advance even when strategic dependence remains asymmetric.

Open at SCMP Ask Claude ChatGPT

China tests a stealth drone from its new amphibious assault ship

# Policy SCMP 18h

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.

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

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

# AI 36Kr 21h

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.

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

China's Middle East strategy faces pressure as Gulf states rearm

# Policy SCMP 11h

Chinese analysis says the prolonged Iran war is pushing Gulf states toward self-reliant defense networks, testing Beijing's strategy of involvement without a binding security commitment.

The article argues that Gulf governments now see neither Washington nor Beijing as able to guarantee protection from Iranian attacks.
China's approach has emphasized diplomatic and economic involvement while avoiding the formal defense obligations carried by the United States; the conflict exposes the limits of that model.
A regional shift toward indigenous defense and diversified partnerships could reduce China's diplomatic leverage even as Gulf states deepen economic ties with Beijing.
The issue is not simply military access: energy security, arms markets, mediation credibility and the safety of trade routes all intersect.
The source presents a strategic dilemma rather than a policy decision by Beijing, so the important question is whether China converts economic influence into more explicit security commitments.

Open at SCMP Ask Claude ChatGPT

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

# Analysis TMTPost 17h

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.

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

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

# AI QbitAI 10h

Alibaba's Qoder desktop client lets users describe software tasks conversationally, then plans, edits, runs and checks the result through browser and computer-use tools.

The product's strongest architectural claim is a closed loop: understand the goal, create a plan, modify a repository, execute it and validate the result in a real environment.
The source describes multi-file changes in an existing Three.js project, which is more relevant to enterprise work than generating a blank demo.
Qoder's agent harness organizes context, files, terminals and tool calls, while browser and computer use provide external validation.
The article says Qoder has served 6 million users and 100,000 companies, but those figures come from promotional material and are not independently substantiated.
The China-specific signal is product positioning: Alibaba is trying to make agentic coding a general desktop workflow for non-programmers while retaining the repository-level capabilities developers need.

Open at QbitAI Ask Claude ChatGPT

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

# Culture TMTPost 17h

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.

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

Japan will deploy fighter jets to India in a China-focused signal

# Policy SCMP 10h

Japan's first planned fighter deployment to India is framed as security cooperation among Indo-Pacific partners responding to China's military rise.

The exercise is scheduled for September 9 to 21 in Rajasthan and is intended to deepen Japanese-Indian defense coordination.
The China dimension is strategic rather than operational: Tokyo is using exercises and defense spending to widen the network around China without creating a formal alliance.
For Beijing, the development complicates the regional environment by linking South Asian and East Asian security partnerships.
The source does not indicate that the deployment changes the balance of air power by itself; its immediate value is signaling and interoperability.

Open at SCMP Ask Claude ChatGPT

HarmonyOS adds on-device semantic image search and super-resolution

# AI InfoQ China 9h

HarmonyOS 7 beta adds natural-language photo retrieval and image super-resolution, giving small Chinese app teams platform primitives instead of a full vision stack to maintain.

The semantic search feature maps natural-language descriptions to indexed images, while super-resolution improves low-quality images after retrieval.
Both capabilities are beta APIs, so device support, build compatibility and real-world output still require testing.
The architectural tradeoff is significant for small teams: the platform can absorb model serving, indexing and device adaptation that would otherwise become a large infrastructure project.
The source stresses complementarity with filenames, tags and OCR rather than treating semantic search as a universal replacement.
The privacy model says images are not retained by the cloud service, but applications still have responsibility for explaining data processing and obtaining consent.

Open at InfoQ China Ask Claude ChatGPT

A Chinese humanoid robot outruns humans but crashes into the finish pad

# AI Solidot 11h

Beijing's Tiangong humanoid reportedly ran 100 meters in 8.64 seconds, revealing impressive actuator output but weak reaction timing, gait efficiency and braking control.

The robot's average speed was nearly 42 kilometers per hour, but it took almost a second to start and more than 50 steps compared with Usain Bolt's 41.
Experts attribute the gait to motor performance rather than human-like sprint mechanics, with shorter steps and higher cadence reducing fall risk.
The robot showed no visible late-race deceleration and collided with the safety pad, underscoring the gap between speed and embodied situational control.
The more meaningful benchmark for general-purpose robots will be turning, perception and autonomous recovery, not a straight-line sprint.

Open at Solidot Ask Claude ChatGPT

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

# Policy TMTPost 17h

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.

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

Intel pitches local AI video production with its 32GB B70 GPU

# AI QbitAI 11h

Intel used a Guangzhou event to promote local AI video workflows for Chinese creators, positioning 32GB of memory and open-model compatibility as alternatives to cloud inference.

The proposed workflow combines local generation, video decoding, super-resolution, asset organization and editing rather than treating the GPU as a standalone accelerator.
The source claims support for a 27B Qwen model with a 256K context window and up to 52 tokens per second, but these are vendor-presented figures.
A hybrid cloud-local model keeps sensitive assets on site while using cloud services for exploration, an increasingly common Chinese enterprise pitch.
The hardware story is adjacent rather than core China news: Intel is foreign, but the event reflects domestic demand for lower-cost, locally controlled AIGC production.

Open at QbitAI Ask Claude ChatGPT

Chinese malls hope 3D printers and robots can restore spending

# Culture TMTPost 14h

Chinese shopping centers are giving prime retail space to 3D printers, robots and AI devices, betting that hands-on novelty can convert rising foot traffic into sales.

Bambu Lab and other hardware companies are opening flagship stores beside luxury and mass-market anchors, while Unitree and robot collectives are using malls for public demonstrations.
Reported store traffic can exceed 10,000 visitors a day, but national mall data show the harder problem is conversion: nearly 30 percent of projects had flat or falling sales and average vacancy reached 9.8 percent.
The products are expensive and still mainly purchased by businesses, while 3D printers can remain unused in homes after the initial novelty fades.
The retail strategy is therefore a market-education experiment as much as a sales channel; malls supply trust and family traffic, while manufacturers test whether spectacle becomes repeat demand.

Open at TMTPost Ask Claude ChatGPT

China Life turns computing and data into operating infrastructure

# Signal TMTPost 9h

China Life says technology is moving from back-office support into product, sales, underwriting and claims, backed by an 8.09 trillion yuan balance sheet and hundreds of deployed agents.

The insurer reports 3.28 billion long-term policies, about 660,000 salespeople and more than 1.75 billion app users, making coordination and data consistency the central engineering problem.
It says a distributed core architecture doubled data-processing capacity, while mixed-cloud infrastructure supports elastic scaling across two locations and three centers.
The company reports more than 90 digital applications and over 500 agents, with intelligent underwriting review reaching 96.8 percent in the first half.
The strategic shift is from isolated automation to a governed enterprise data and model layer that can be reused across a national service network.
The claims are corporate disclosures, so the important question is whether these systems improve loss ratios, service quality and human workload rather than merely increasing reported automation rates.

Open at TMTPost Ask Claude ChatGPT

Xiaomi foldable phone to debut CXMT LPDDR6 memory

# Signal Techmeme 17h

A Techmeme roundup says Xiaomi's next foldable will pair CXMT LPDDR6 memory with Xiaomi's in-house 3nm processor, a potential Chinese silicon stack milestone.

The report attributes the supply announcement to Xiaomi and CXMT and places the launch in September.
If confirmed in a shipping flagship, the combination would show domestic memory, application silicon and device integration moving together rather than advancing as isolated components.
The supplied item does not provide memory speed, capacity, yields or independent confirmation of the processor's manufacturing performance.
Because this appears through an aggregator, the claim should be treated as a reported milestone pending the product launch.

Open at Techmeme Ask Claude ChatGPT

Yunnan revokes tour operator's license after coercive-sales case

# Policy Jiemian 8h

Yunnan authorities say a tour guide threatened travelers who rejected paid excursions, revoking the agency's license and detaining the guide.

Officials say the two viral videos concerned the same tour group and found deception and coercion in the guide's sales practices.
The response included license revocation for the agency, cancellation of the guide's credential and administrative detention.
The case shows how viral consumer complaints can trigger rapid local enforcement when they expose a persistent tourism-market abuse.

Open at Jiemian Ask Claude ChatGPT

Huawei pitches a hybrid data layer for enterprise agents

# AI Leiphone 15h

Huawei Cloud says enterprise data platforms must evolve from storage and reporting toward governed feeds for model training, agent execution and trusted data exchange.

The proposed stack combines hybrid cloud, multimodal data lakes, metadata management, data-processing operators and privacy-preserving exchange.
Examples include a Guangdong utility's labeled data system and a regional data network covering 232 institutions and 40 cities, but these are vendor-presented deployments rather than independent evaluations.
The strategic message is that Chinese enterprises want AI access to operational data without moving all sensitive records into a public cloud.

Open at Leiphone Ask Claude ChatGPT

Chabaidao expands from tea drinks into all-day consumption

# Signal 36Kr 12h

Chabaidao reported 2.655 billion yuan in first-half revenue, roughly $374 million, as a fast-selling lychee milk drink and a broader store network pushed the chain beyond afternoon tea.

The company says its lychee milk drink sold more than 17 million cups in about 50 days, while revenue rose 6.2 percent year on year.
It expanded into breakfast, light meals, coffee and hydration products, using 27 distribution centers and 91 percent fruit allocation to support the range.
The chain had 8,863 stores and more than 200 million registered members, but those figures do not establish customer frequency or profitability by new category.

Open at 36Kr Ask Claude ChatGPT

A Chinese game and XR glasses turn fandom into hardware

# Culture Zhidx 19h

VITURE's limited-edition XR glasses tie a popular Chinese martial-arts game to a portable giant screen, showing how domestic entertainment IP is becoming a hardware design and retail asset.

The glasses are based on VITURE Beast and offer a reported 58-degree field of view with a virtual screen up to 174 inches at four meters.
The collaboration extends beyond cosmetics into packaging, startup menus, interface themes and sound effects, making the device a collectible as well as a display.
The game reportedly sold about 300,000 preorders in its first week and generated more than 100 million yuan, roughly $14 million, in revenue.
The product is still a premium accessory launch rather than evidence that XR has become a mass gaming platform.

Open at Zhidx Ask Claude ChatGPT

AI-made dramas take over China's short-video charts

# Culture Techmeme 12h

The Financial Times, surfaced by Techmeme, says AI productions made up 89 of the 100 most popular animated dramas on Douyin in May, putting online performers under pressure.

This is an aggregator item, so it shows what US technology media is noticing rather than providing original verification.
The China-specific signal is labor displacement inside a fast-moving short-video economy, where cheaper generation can change both production volume and the value of human performers.
The figure is attributed to DataEye and describes a particular chart and month, not the whole Chinese entertainment market.

Open at Techmeme Ask Claude ChatGPT

ChinAI	403	0/0		HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk	200	0/20
Pekingnology	200	0/20
Ginger River Review	200	0/20
Sinocism	200	0/20
QbitAI	200	7/10
36Kr	200	3/30
Solidot	200	5/20
Leiphone	200	2/20
GeekPark	200	2/30
IT Home	200	12/60	+39 over cap
SCMP	200	10/50	+4 over cap
Initium	200	1/15
Jiemian	200	8/30	+22 over cap
China Digital Times	200	0/15
The Wire China	200	0/50
Techmeme	200	5/15
Hacker News	200	2/20
V2EX	200	8/50	+8 over cap
Zhidx	200	1/20
InfoQ China	200	2/20
New Wisdom	200	8/15	+1 over cap
TMTPost	200	8/19	+11 over cap

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

CXMT takes Section 1260H to a Washington judge

Xi lands in Bishkek before he lands in Washington

Alibaba went sparse, Meitu went to the bank

A robot made 55-step ice cream in Shanghai

A programmer died in a company toilet

Threads we are pulling