China Desk

China News, Summarized 13 Aug 2026, 23:01 UTC 34 stories 18/19 sources

Today's brief

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

The model landed first, quietly. At around 3am Beijing time the API docs flipped: the fingerprint behind deepseek-v4-pro became the 0813 build, 111 days after the preview, with no blog post and no changelog. Architecture untouched — 1.6 trillion total parameters, 49 billion activated, 1M context, 384K max output. Everything that moved, moved in post-training.

And it moved a lot. DeepSWE, DeepSeek's own software-engineering benchmark, went from 12.8 to 62.7. Cybergym 52.7 to 83.3. Terminal Bench 72.1 to 87.9. Those are all long-horizon code, terminal and tool-calling tasks — exactly where the preview fell over. On pure knowledge reasoning it still trails: HLE 42.7 without tools against Claude Opus 4.8's 49.8 and Fable 5's 53.3.

Then came the invoice. DeepSeek is introducing peak and off-peak billing, off-peak at half the peak rate, new prices live at 16:00 UTC on 16 August. Reuters read the statement as rates ranging from 50% to 1,100% above current prices depending on model, token type and hour. A developer's tabulation on Hacker News put V4-Flash peak output at $1.32 per million against $0.28 today.

Look at when peak is. 01:00–04:00 and 06:00–10:00 UTC — the Chinese working day, morning and afternoon. Cache reads take the worst multiple, and cache reads are what agent loops burn. That is not margin expansion; that is load-shedding by price, from a lab that started hiring electrical and HVAC engineers for its own data centres two weeks ago.

Why now is in the leaked minutes. ChinaTalk has been working through notes from a four-hour meeting between Liang Wenfeng — the quant fund manager who founded DeepSeek and is now, on paper, richer than Dario Amodei or Sam Altman — and his investors. Enterprise revenue this year in the hundreds of millions of dollars. Net profit soon. Then an IPO.

A company priced for a listing cannot subsidise the planet's inference. The February 2025 disclosure that R1's API earned $562,027 a day at a 545% margin was a flex; it is now a floor to defend. Developers on V2EX spent the day joking that the rheostat needs a new setting and complaining that models were shipped and pulled, a mess. The cheap era was a customer-acquisition budget, and it has been spent.

What Beijing is signalling

Washington published 25 pages instead. The White House Office of Trade and Manufacturing Policy released "The Great Transshipment Scam", alleging Chinese exporters have laundered origin through more than 40 third countries since 2018 via relabelling, re-invoicing and minor assembly. It estimates the annual value at $40bn to $303bn and, assuming $75bn, claims 450,000 displaced American jobs.

The targets are the middlemen. Korea, Japan, Taiwan, the EU, India, Mexico and Canada are filed as "diversified scale leaders"; Vietnam, Malaysia, Thailand, Turkey, Brazil and Indonesia as deeply China-integrated ones. Nothing in today's mainland feed touches it — the Chinese press is on Zhu Rongji and V4-Pro. Xi arrives in Washington next month with this on the table and the investment board still unbuilt.

Nuclear is the AI answer nobody can sanction. CCTV Finance puts 57 reactors under construction and 135 GW operating-plus-approved, the largest build programme in the world. Eight units were approved in late July — the first batch of the 15th Five-Year Plan. Nuclear is under 2% of installed capacity but nearly 5% of generation, and above 15% in Guangdong, Fujian, Liaoning and Hainan.

Set that against yesterday's number: 800 billion kWh of compute electricity demand by 2030. Reactors approved now come online in the 2030s. Beijing is answering a question about tokens with twenty-year concrete.

And state TV is now bragging about exploit agents. A Fudan University team's Whitzard agent scored 91.2% on CyberGym, second globally and first among universities, ahead of Anthropic's security-focused frontier model and teams from Microsoft and Google, across 1,507 real vulnerabilities in 188 large open-source projects. One base model, path reachability recast as reasoning constraints, total inference cost under 5,000 yuan (~$700). Framework and benchmark open-sourced, and CCTV News carried it.

Go deeper on this section: ClaudeChatGPT

The AI race

Tencent's free cash flow went negative, on purpose. Q2 revenue 204.79bn yuan (~$28.8bn), up 11%; capex 52.78bn yuan (~$7.4bn), up 176% year on year and 65% on the quarter; R&D up 35%. The market hated it — shares closed down 4.46% the next morning.

The interesting claim is what the money buys. 36Kr's read is that Tencent is hoarding high-spec memory as much as accelerators, with Samsung, SK Hynix and Micron effectively sold out and expansion running one to three years behind demand. Hunyuan's next generation is meant to be bigger. Racing the clock on DRAM is a stranger bet than racing it on GPUs, and a more honest one.

DeepSeek shipped Harness the same day. A developer preview, one npm install, everything a plugin, built on an existing open-source plugin framework. What US developers on Hacker News fixed on was the raw reasoning trace being visible where American labs obfuscate theirs.

The leaked shortlist is the real strategy. Of 17 accepted ecosystem projects, about 70% are near-zero-star personal repos — sandboxes with credential injection at the gateway, cost-aware model routing, persistent "failure notebooks", multi-agent schedulers. Famous high-star projects were rejected; medical, legal and financial verticals are under 4%. DeepSeek is buying the execution layer and leaving applications to everyone else.

Meanwhile, the least glamorous work of the week. The FlagOS community got Alibaba's 2.4-trillion-parameter Qwen3.8 running day-zero on nine different accelerator families — Nvidia plus Ascend, Moore Threads, Kunlunxin, Hygon, MetaX, Enflame and others — largely by adding INT8 quantisation paths, because more than 80% of China's installed AI capacity cannot do FP8. A frontier model reaching hardware bought two generations ago is worth more than a leaderboard position.

Go deeper on this section: ClaudeChatGPT

The vibe

The Unitree lottery is the story of the week in China, and it isn't about robots. The humanoid maker's STAR Market allotment rate finished at 0.01809759%, a record low, with more than 9.78 million accounts applying for 19,414 winning numbers of 500 shares each. One Chengdu winner told reporters he expects about 200,000 yuan (~$28,000) and will pay cash for a Xiaomi car.

The detail that lands is what he won't do. He won't post it to WeChat Moments — afraid colleagues will resent him and relatives will ask for loans. Also in the book: Liang Wenfeng's two funds took the largest private-fund allocation by account count, 198 products for about 38.94m yuan (~$5.5m).

Someone in Hangzhou built a flying skateboard. Kufei launched a standing single-person eVTOL — no seat, no cabin, a flattened black platform with caged rotors, about 11 minutes aloft with a 70kg pilot — plus a seated model doing 100 km/h for 40 minutes, on a dual-redundant in-house flight controller with three aviation-grade IMUs per channel. It flew live at the Spring Festival Gala's Harbin stage at minus 30.

The route to market is American. Roughly 100 paid deposits overseas, aimed at FAA Part 103 ultralights, which need no airworthiness certificate and no pilot licence — plus a package from PICC and CPIC covering aircraft, occupant and third parties. A Chinese consumer aircraft finds its first legal home in a 1982 US rule.

Go deeper on this section: ClaudeChatGPT

Read differently

One release, two entirely different stories. Chinese tech press led on benchmarks and explicitly reported the price as unchanged at 3 yuan in, 6 yuan out per million (~$0.42/$0.85) — QbitAI called it a rehearsal of the 2025 Spring Festival. English-language developer forums skipped the benchmarks and went straight to the hike, doing substitution arithmetic against OpenAI's Luna price cut and noting DeepSeek is still roughly twice as slow.

Only the Chinese write-ups tested it properly. GeekPark found thinking mode consuming 13,394 of 16,384 output tokens on a pathfinding task and truncating the HTML mid-file; with thinking off, 715 clean lines in 54 seconds. The chain of thought itself has degenerated into clipped telegraphese to save tokens. Great for agents, probably fatal for prose.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Called it, and fast. Yesterday's note on a registered Harness account and data-centre job ads; the developer preview shipped within the day.
  • Still missing: Alibaba's promised 27B dense companion to Qwen3.8. The multi-chip port covers only the 2.4T text checkpoint.
  • Lin Junyang's new lab is confirmed as an angel round at a $2bn valuation, co-led by Gaorong and HSG with Tencent and a Shanghai future-industries fund.
  • Zhu Rongji: tributes, including a personal one from Malaysia's Anwar Ibrahim, but no funeral grade yet. Who eulogises is still the thing to watch.
  • Insta360 outgrew its own category: revenue approaching 10bn yuan against a global panoramic market Leiphone puts under 7bn. Which explains the pivot.

Go deeper on this section: ClaudeChatGPT

The river · 34

Zhu Rongji's centralized reform legacy

# Initium 23h

Chinese coverage presents Zhu Rongji's death as a reckoning with the economic centralization that modernized China, integrated it into global trade and left workers with lasting collective trauma.

Also: SCMP, The Wire China

More
  • The account says Zhu died in Beijing on August 12 at age 97 and recalls his public image as an incorruptible premier willing to confront officials and vested interests.
  • His reforms strengthened central control over taxation and economic policy while accelerating market restructuring, state-enterprise closures and global integration.
  • The result was a more coherent national fiscal system and the platform for export growth, but also a generation of laid-off industrial workers and households that experienced reform as dispossession.
  • The phrase economic czar captures both sides of the legacy: he was an unusually forceful administrator, and his concentration of authority made large-scale change possible without broad public consent.
  • The story matters for understanding today's China because many current fiscal, regional and labor problems are downstream of the reform architecture he helped build.
  • Chinese memory of Zhu is unusually layered: official praise, popular admiration for his bluntness and private resentment over social costs can coexist without producing a single consensus.
  • His death removes one of the last senior figures associated with the reform era and invites comparison with a present political economy that is more security-centered and less openly market-oriented.
Open at Initium Ask Claude ChatGPT

DeepSeek V4 Pro formalizes a cheap, long-context agent model

# GeekPark 14h

DeepSeek quietly replaced its preview with a 1.6-trillion-parameter MoE model whose agent benchmarks jumped sharply, while real testing shows both stronger tool use and a costly thinking-mode failure.

More
  • The architecture reportedly remains 1.6 trillion total parameters with 49 billion active, a 1 million-token context window and 384,000-token maximum output; the major change is post-training.
  • The cited scores rose from 12.8 to 62.7 on DeepSWE, from 52.7 to 83.3 on CyberGym and from 72.1 to 87.9 on Terminal Bench, all tasks involving code changes, environment operations or tool calls.
  • The API price remains 3 yuan per million input tokens, roughly $0.42, and 6 yuan per million output tokens, roughly $0.85, putting it far below leading closed models.
  • Hands-on tests produced complete interactive web applications, but a complex pathfinding task consumed most of its output budget on internal reasoning and truncated the actual code until thinking mode was disabled.
  • That behavior exposes a key systems issue: maximizing reasoning tokens can reduce useful delivered work when output limits are fixed, so model settings and harness policy become part of application design.
  • The model still trails cited competitors on some knowledge and software-engineering tests, meaning its advantage is price-performance and agent usability rather than universal capability leadership.
  • DeepSeek's use of its own Harness in agent evaluations signals that the company is optimizing a model-runtime pair, not treating the base model as a standalone commodity.
  • For Chinese developers, the release also shows how a low-cost API can accelerate domestic adoption even as planned price increases create urgency around caching and deployment choices.
Open at GeekPark Ask Claude ChatGPT

Qwen's 2.4-trillion-parameter model runs on nine chip ecosystems on day one

# Leiphone 14h

Alibaba open-sourced Qwen3.8-2.4T-A95B, while China's FlagOS community completed same-day deployment work across nine AI-chip platforms to make existing hardware usable for the model.

More
  • The mixture-of-experts model has 2.4 trillion total parameters and 95 billion active parameters, with a native 262,144-token context expandable to about 1 million tokens.
  • FlagOS says it validated BF16, FP8 and INT8 versions across chips from nine Chinese and foreign vendors, using a common compiler, operator and plugin stack rather than bespoke model ports.
  • The engineering challenge was scale: the model is roughly six times larger than the cited 397-billion-parameter predecessor, forcing work on quantization, memory pressure, parallelism and inter-chip communication.
  • The project supports W8A8 INT8 quantization to reduce weight and activation traffic, potentially allowing older deployed accelerators to serve newer models instead of becoming stranded assets.
  • This is strategically important for China because software portability can partially offset hardware-access constraints and reduce dependence on one dominant accelerator stack.
  • The claimed sub-24-hour adaptation and accuracy alignment are ecosystem claims from the project; independent workload testing will determine whether day-one availability translates into useful throughput.
  • If the common stack works, the moat shifts from chip peak performance toward compilers, kernels, deployment tooling and the ability to keep installed capacity productive.
Open at Leiphone Ask Claude ChatGPT

Hong Kong's July courts show national-security law's widening reach

# Initium 16h

Hong Kong authorities raided independent bookstores, convicted an autistic young man of sedition and forced a judge into early retirement, while a visa challenge by mainland students failed.

More
  • National-security police arrested five bookstore workers and seized books under the city's Article 23 law, amid a broader pattern of independent bookshops closing because the political red lines are unpredictable.
  • An 18-year-old with autism was convicted over slogans written in a public bathroom and later received an 18-month probation order, the first non-custodial sentence under the new sedition provision.
  • The judiciary also required a judge to retire early after repeated findings of copied reasoning, an episode that highlights institutional accountability but also exposes a court system under unusual public scrutiny.
  • A separate case involving 87 mainland students challenged restrictions on access to subsidized university places for dependent-visa holders; the court upheld prioritizing applicants with a closer local connection.
  • Together the cases show two directions of governance: tighter control over speech and civil society, alongside administrative and judicial rules designed to allocate scarce public resources.
  • The bookstore raids are especially significant for information flows because they target distribution and possession of texts, not merely public political organizing.
  • Hong Kong's legal environment is therefore becoming a mechanism for both national-security enforcement and social boundary-setting, with consequences for media, education and cultural commerce.
Open at Initium Ask Claude ChatGPT

Tencent's AI spending pushes quarterly free cash flow negative

# 36Kr 13h

Tencent's second-quarter capital expenditure rose 176 percent year on year to about $74.3 billion yuan equivalent? Chinese analysis says 52.8 billion yuan, roughly $7.4 billion, as it races to secure compute and memory for AI.

More
  • Tencent reported 204.8 billion yuan in quarterly revenue, roughly $28.8 billion, and 67.3 billion yuan in operating profit, roughly $9.5 billion; adjusted operating profit excluding newer AI products was higher.
  • Research spending reached 27.3 billion yuan, about $3.8 billion, while capital expenditure reached 52.8 billion yuan, about $7.4 billion, making free cash flow negative for the quarter.
  • The analysis argues that the spending is not only a GPU purchase: large models also require high-bandwidth memory and storage, whose supply expands more slowly than AI demand.
  • Tencent is using mature games and advertising as cash engines for a bet on models, CodeBuddy, WorkBuddy and other AI services, accepting near-term margin pressure for strategic capacity.
  • The market reaction, including a reported 4.46 percent share-price fall, shows that investors want evidence of monetization before accepting Silicon Valley-style infrastructure spending.
  • The strategic risk is inventory and utilization: if model demand or product adoption disappoints, expensive memory and compute can become a drag rather than a moat.
  • The story matters because it captures China's platform companies moving from cautious AI experimentation to balance-sheet-scale infrastructure commitments.
Open at 36Kr Ask Claude ChatGPT

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

# AI QbitAI 6h

Chinese testing finds DeepSeek Harness unusually configurable, with pluggable models, tools, memory and context management; US discussion instead probes its runtime design and multi-provider potential.

Also: Leiphone, Hacker News, V2EX

More
  • The product is open source and exposes the components that many agent products keep proprietary: model selection, tools, policies, storage and context handling.
  • Its trajectory view records raw event-level execution rather than only a polished chat transcript, giving developers visibility into tool calls, failures and token costs.
  • The report says more than 100 plugins are included and that the system can use other model providers, which makes it closer to an agent operating layer than a single-model application.
  • A notable design choice is active clarification: when a request is underspecified, the harness asks questions and proposes options instead of confidently producing a partial result.
  • The author reports that long-running tasks improved substantially compared with a competing coding agent, but this is a personal test rather than a controlled benchmark.
  • The tradeoff is product friction: there is no native desktop app, the interface is browser-based, and several workflow panels remain less mature than established competitors.
  • Chinese coverage frames the release as an eccentric developer saddle; US discussion treats it as a composable runtime whose value may lie in routing work across different models.
Open at QbitAI Ask Claude ChatGPT

Claude helps construct a 668-order Hadamard matrix

# AI QbitAI 11h

Chinese coverage says Anthropic researcher Levent Alpoge and Claude produced a verified 668-order Hadamard matrix, potentially resolving the smallest long-standing open case below order 2,000.

More
  • A Hadamard matrix contains only plus-one and minus-one entries, with every pair of rows having zero inner product; the 668-order case had resisted construction for decades.
  • The reported solution includes 12 matrices covering every previously missing order below 2,000, and independent exact-integer checks reportedly verified the orthogonality conditions.
  • The result is not yet a general proof of the Hadamard conjecture: the benchmark organization has provisionally marked the problem as AI-solved while awaiting a technical account.
  • The research workflow is human-machine collaboration, with the model helping decode and explore a construction that humans then checked precisely.
  • If the method generalizes, it would matter to combinatorial design and potentially coding and signal-processing theory; if it is only a one-off search artifact, its broader significance is smaller.
  • The story is a better test of agentic mathematical work than a simple answer-generation benchmark because the output is a large, checkable object rather than a persuasive explanation.
Open at QbitAI Ask Claude ChatGPT

Lenovo's AI infrastructure revenue nearly doubles in record quarter

# Signal QbitAI 16h

Lenovo reported 183.4 billion yuan in quarterly revenue, roughly $25.8 billion, while infrastructure revenue nearly doubled and AI-related sales reached about $89.3 billion yuan equivalent?

More
  • The company reported 63.4 billion yuan in AI-related revenue, roughly $8.9 billion, or about 35 percent of group sales, and adjusted net profit above $1 billion.
  • Infrastructure Solutions revenue reached 57.9 billion yuan, roughly $8.2 billion, with a 9.1 percent operating margin, making the division a new profit engine rather than a low-margin server adjunct.
  • The numbers show Lenovo benefiting from hyperscaler AI infrastructure spending while using its global supply chain and local delivery model to manage component inflation and geopolitical fragmentation.
  • Lenovo says its personal AI agent has a 40 percent weekly active rate on PCs and that Qira converts about 28 percent of users, but these are company metrics without methodology.
  • The strategic shift is from selling devices into supplying hybrid AI systems, services and infrastructure, with AI demand now linking Lenovo's client and data-center businesses.
  • That model gives Lenovo more leverage than a pure PC maker, but also exposes it to capital-cycle swings and the risk that AI infrastructure margins normalize as supply catches up.
Open at QbitAI Ask Claude ChatGPT

The DeepSeek thesis: open models, constrained hardware and a long horizon

# Analysis ChinaTalk 12h

ChinaTalk argues that DeepSeek founder Liang Wenfeng sees automated learning and general intelligence as a long-term mission, using China's hardware constraints as an engineering condition rather than a strategic retreat.

More
  • The analysis is based on leaked meeting minutes and the author's interpretation, not a formal statement of company strategy.
  • It portrays DeepSeek's open-model economics as relying on business customers and ecosystem adoption rather than selling model weights directly.
  • The reported worldview helps explain why the lab repeatedly releases capable systems at low prices: distribution and technical iteration may be more valuable than near-term licensing revenue.
  • The Chinese political-economy angle is important because the lab is described as unusually insulated from ordinary commercial and political trend-following within China's technology sector.
  • The account also compares Liang with Western lab leaders, but the strongest China-specific insight is institutional: a privately backed Hangzhou group can pursue a high-risk research thesis while working around constrained access to advanced hardware.
  • The interpretation remains provisional until the leaked minutes, management structure and revenue model are independently confirmed.
Open at ChinaTalk Ask Claude ChatGPT

Hangzhou startup unveils a standing personal eVTOL

# AI QbitAI 12h

Chinese startup Coolfly unveiled standing and seated single-person electric aircraft, including an 11-minute standing model with redundant flight control and nearly 100 paid overseas deposits.

More
  • The standing Urban model must compensate in real time for the pilot's shifting center of gravity, making the human-aircraft combination a dynamic control problem rather than a simple multirotor platform.
  • Its NA80 flight-control system uses two independent computers and three aviation-grade inertial sensors in each, with claimed millisecond fault detection and failover.
  • The seated Dream Pro is reported to carry 120 kilograms, reach 100 kilometers per hour and fly for up to 40 minutes with a 70-kilogram pilot.
  • The company has arranged a bundled insurance product and is targeting markets with ultralight-aircraft rules, which may lower certification and pilot-licensing barriers.
  • The commercialization signal is real but narrow: paid deposits and permissive rules do not yet demonstrate safe, scalable personal aviation in dense urban environments.
  • China's low-altitude-economy push is creating a pathway from public demonstrations to consumer products, but regulation, training and liability remain the limiting systems.
Open at QbitAI Ask Claude ChatGPT

SCALEFORCE raises several million yuan for embodied-AI data infrastructure

# AI QbitAI 11h

Chinese startup SCALEFORCE raised several tens of millions of yuan, roughly $4-7 million, to build MatrixOS and a physical-world data network for embodied AI.

More
  • The company argues that robots need synchronized visual, tactile, joint-trajectory and physical-interaction data that cannot be scraped from the internet like text.
  • It claims the available pool of high-quality physical-interaction data is about 500,000 hours against a requirement of tens of millions of hours, although those market estimates are not independently substantiated here.
  • MatrixOS is positioned as a full data loop from collection and processing to customer deployment, with a neutral infrastructure role across robot makers and world-model companies.
  • The startup says it has a million-yuan-scale data order from a robotics customer and partnerships with Chinese universities, Ascend's ecosystem and a state-media robotics initiative.
  • The strategic question is whether data collection networks develop durable quality and distribution advantages, or become labor-intensive services with weak margins once robot makers build internal pipelines.
Open at QbitAI Ask Claude ChatGPT

US accuses Chinese exporters of a $40-country transshipment scheme

# Policy SCMP 12h

Chinese coverage leads with Washington's allegation that Chinese exporters routed goods through more than 40 countries to evade tariffs, while the policy issue is a new escalation in customs enforcement.

More
  • The 25-page US report calls the alleged practice the Great Transshipment Scam and says it involves relabelling or limited assembly in lower-tariff jurisdictions.
  • The accusation extends the trade conflict from tariffs on Chinese goods to scrutiny of supply-chain geography, documentation and the substance of processing in third countries.
  • For Chinese manufacturers, the implication is that nominal origin rules and contract-manufacturing structures may no longer shield exports from US enforcement.
  • The report is an allegation by the US administration, not a court finding; the supplied account does not identify specific companies or quantify the alleged trade flows.
  • If Washington turns the report into seizures, penalties or new origin rules, multinational supply chains will face higher compliance costs and more pressure to localize final production.
  • Chinese coverage's emphasis on the US accusation, rather than on a confirmed enforcement action, signals that the immediate event is political positioning as much as customs policy.
Open at SCMP Ask Claude ChatGPT

iFlytek launches enterprise agents across seven business functions

# AI QbitAI 12h

Chinese AI company iFlytek moved from isolated enterprise tools toward a unified agent stack covering management, IT, procurement, sales, manufacturing, analytics and overseas operations.

More
  • The platform combines a common agent foundation with domain workflows, reflecting a shift from generating documents to executing multi-step business processes with traceable outputs.
  • Its procurement tools are designed to analyze projects continuously across people, companies, prices, behavior and outcomes, while compliance review follows a chain from rule interpretation to evidence retrieval and human review.
  • The manufacturing offering includes Agentic RAG, voice-model training and an appliance-style Agent OS intended to run on enterprise hardware rather than only in the cloud.
  • The company says its speech stack supports more than 100 languages for recognition and more than 120 for synthesis, with service nodes in five overseas regions.
  • The strategic challenge is governance: enterprise agents need permissions, auditability and domain validation, not just a capable model.
  • The launch reflects a broader Chinese enterprise-AI pattern in which vendors sell an integrated stack of model, workflow, hardware and deployment services to institutions that may not have large internal engineering teams.
Open at QbitAI Ask Claude ChatGPT

A black-box attack makes vision-language navigation robots lose their way

# AI Leiphone 13h

A Hong Kong research team developed AdvNav, a black-box adversarial attack that uses observed behavior to steer vision-language navigation agents away from their targets.

More
  • AdvNav does not require model parameters or gradients; it perturbs the robot's visual input and uses trajectory-level and action-level feedback to optimize the disturbance.
  • The method combines adaptive perturbation strength with a genetic algorithm, addressing the delayed and sparse feedback that makes attacks on sequential navigation harder than ordinary image classification attacks.
  • On the cited R2R benchmark, the attack success rates against three systems ranged from 49.70 percent to 87.30 percent, but the supplied text does not map each rate cleanly to a named model.
  • The result matters for deployed robots because a small visual change can compound across a route, producing a large navigation failure even when each individual action looks plausible.
  • Black-box testing is particularly relevant to commercial systems whose internal weights and gradients are inaccessible to customers or security researchers.
  • The missing deployment evidence is whether the perturbations survive real cameras, lighting, motion and physical-world defenses rather than only benchmark images.
Open at Leiphone Ask Claude ChatGPT

Grok 4.6 targets long-running agents with lower pricing

# AI QbitAI 10h

Chinese coverage says Grok 4.6 leads several knowledge-work benchmarks at $2 per million input tokens and $6 per million output tokens, while remaining weaker on terminal-only tasks.

More
  • The reported model scores 1,753 on GDPVal-AA v2, ahead of the cited GPT and Fable models, but reaches only 26 percent on Terminal-Bench versus 34.6 percent for GPT.
  • Its intended workload is a long-running cloud agent that can use websites and tools, pause for human approval and resume without the user's computer remaining open.
  • The pricing is attractive but doubles for requests over 200,000 tokens and doubles again for a low-latency version, so workload shape matters more than the headline rate.
  • The Chinese account places Grok alongside domestic workplace agents such as WorkBuddy, Trae and QoderWork, framing the contest as a battle to occupy the employee's workflow rather than merely win model benchmarks.
  • The key unresolved issue is reliability over unattended, multi-step execution: the supplied benchmarks show strength in knowledge work but not a universal lead.
Open at QbitAI Ask Claude ChatGPT

Didi reports 5.05 billion quarterly platform orders

# Signal Leiphone 12h

Didi says core-platform orders rose 13.2 percent to 5.05 billion in the second quarter, with China mobility growing for a 14th straight quarter and total transaction value reaching about $188.6 billion.

More
  • China mobility produced 3.648 billion orders and 904 billion yuan in gross transaction value, roughly $127 billion, according to the company.
  • Didi says demand is shifting from merely finding a car toward predictable vehicle quality and service, with a premium ride category reporting a 98 percent approval rate and lower complaints.
  • An AI service layer with more than 90 user-intent labels and a multi-agent matching architecture reportedly lifted satisfaction in core scenarios by nearly ten percentage points.
  • The company also links more than 74,000 charging stations across over 290 Chinese cities and says its international orders rose 29 percent to 1.403 billion.
  • The numbers show Didi evolving from a ride-hailing marketplace into a service and infrastructure platform, while still relying on high-frequency mobility as its anchor.
  • The main strategic uncertainty is whether international growth and AI-assisted service quality can offset the mature economics and regulatory exposure of China's core market.
Open at Leiphone Ask Claude ChatGPT

China's electric-car exports increasingly travel through local operating networks

# Analysis Leiphone 13h

Didi's Brazilian platform 99 and rental company Movida will offer cheaper vehicle leases to as many as 50,000 drivers, showing how Chinese mobility know-how can help localize exported EVs.

More
  • The planned program is designed to cut rental costs by about 25 percent and includes electric vehicles, addressing the gap between lower operating costs and higher upfront access costs for Brazilian ride-hailing drivers.
  • Didi is using 99's local network of users, drivers, payments and delivery services as distribution infrastructure rather than treating vehicle exports as a standalone car sale.
  • The report says a Brazil sustainable-mobility alliance now includes 31 companies and has attracted 410 million Brazilian reais of investment, described as about 536 million yuan or roughly $75 million.
  • The model resembles China's own EV adoption path, where ride-hailing platforms, financing, charging and high-mileage drivers helped turn electric cars into everyday production tools.
  • This is a substantive China story because it shows Chinese firms exporting an operating ecosystem and local-market playbook, not only hardware.
  • The unresolved issue is whether the network can create durable local value or whether it will be seen as another channel for Chinese industrial overcapacity.
Open at Leiphone Ask Claude ChatGPT

Chinese STEM students turn to Singapore as Western study routes narrow

# Culture SCMP 1h

Chinese science and engineering students are increasingly considering Singapore as US, UK, Canadian and Australian immigration and technology restrictions shrink the old Westward path.

More
  • The shift reflects a structural change in China's talent pipeline: university choice is increasingly shaped by visa access and geopolitical risk, not only academic prestige.
  • Singapore offers geographic proximity and a comparatively open technology environment, while preserving access to regional research and industry networks.
  • If the pattern persists, Chinese companies and universities may benefit from a larger regional talent network even as Western institutions lose some of their traditional recruiting advantage.
Open at SCMP Ask Claude ChatGPT

Robot developers may be underestimating the hardware body

# Analysis Leiphone 17h

Chinese robotics coverage argues that embodied-AI progress is constrained not only by models and data but by robot bodies that lack low latency, precision, endurance and open interfaces.

More
  • A robot body converts model outputs into physical action; motor jitter, thermal drift and poor control interfaces can erase the value of a sophisticated policy.
  • Stable hardware is also a data requirement because useful training records need synchronized vision, joint state, force and audio signals over long continuous runs.
  • The report presents a competition-winning TA2 platform as an example of a reusable body with one control interface for teleoperation and model inference, reducing the gap between data collection and deployment.
  • It says the robot synchronizes three visual streams with full state data and has operated in an open retail environment after multiple proof-of-concept rounds.
  • The broader market implication is that a well-documented, tested robot body can become shared infrastructure for many model teams, much as a cloud platform accelerates software experimentation.
  • This challenges the industry's model-first narrative: the scarce asset may be a reliable physical interface and a high-quality execution loop, not another general-purpose policy model.
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 experts caution that divergent interests make a NATO-style alliance unlikely.

More
  • The agreement promises collective defence, treating an attack on one signatory as an attack on all, and links Saudi capital, Turkish military capacity and Pakistani nuclear status.
  • A second Chinese account stresses the gap between political symbolism and operational capacity: the members have different strategic priorities and limited mechanisms for mutual defence.
  • For Beijing, the pact matters less as a new alliance than as evidence that regional states are exploring security arrangements less dependent on Washington.
  • The agreement could create more room for Chinese diplomacy and economic engagement, but that would require the three states to coordinate beyond a broad deterrence statement.
  • Its practical significance will depend on whether the parties build command, logistics and intelligence mechanisms rather than merely repeat the pledge.
Open at SCMP Ask Claude ChatGPT

Unitree's IPO draw shows China's robot-stock frenzy

# Signal Leiphone 22h

Chinese coverage pairs Unitree's record-low 0.018 percent IPO lottery rate with DeepSeek V4 Pro's release and a US reversal on TikTok device restrictions, revealing a market hungry for AI and robotics exposure.

More
  • More than 9.78 million investors reportedly applied for Unitree shares, with 19,414 winning numbers and 500 shares available per number; the lottery rate was described as the lowest in the exchange's history.
  • Forum reactions include winners hiding the result from friends and estimating gains of about 200,000 yuan, roughly $28,000, but these are personal expectations rather than evidence of future returns.
  • The Unitree demand signal matters because it shows how China's public market is treating robotics as a mass-investor theme, not merely an industrial technology sector.
  • The roundup also reports DeepSeek V4 Pro at 3 yuan per million input tokens, roughly $0.42, and 6 yuan per million output tokens, roughly $0.85, reinforcing the domestic narrative of cheap frontier AI.
  • The item is a roundup, so individual claims are relayed from multiple outlets and should not be read as a single verified account.
  • Together the stories show China's AI economy operating through both engineering adoption and financial-market enthusiasm, with a risk that public excitement outruns deployment economics.
Open at Leiphone Ask Claude ChatGPT

Chinese missile AI reportedly classifies stealth-fighter heat signatures

# Analysis SCMP 9h

SCMP reports a Chinese system that identifies F-22- and F-35-like infrared signatures with more than 90 percent accuracy, though the supplied excerpt lacks testing details.

More
  • The reported approach exploits the fact that stealth reduces radar visibility but cannot eliminate engine exhaust and aerodynamic heat.
  • A system that distinguishes aircraft heat from decoy flares could improve missile seekers or broader infrared target recognition, but accuracy depends heavily on sensor range, weather and adversarial countermeasures.
  • The source excerpt does not identify the laboratory, dataset or whether the figure comes from simulation, field tests or a military exercise.
Open at SCMP Ask Claude ChatGPT

Chinese AI research heads to IJCAI amid a shift from scale to reasoning

# Analysis Leiphone 13h

Chinese coverage says IJCAI 2026 is drawing strong Chinese participation despite a domestic rating downgrade, with industry research moving toward planning, symbolic reasoning, safety and robustness.

More
  • The report says the conference received more than 6,500 valid submissions and accepted 990 papers, while Chinese institutions remained a major contributor.
  • It claims reasoning, planning, constraint solving and search together account for about 20.8 percent of main-track papers, though the statistics are presented by the outlet rather than independently verified here.
  • The framing contrasts sponsor visibility with technical participation: Chinese companies may spend less on logos while publishing work tied to fraud detection, industrial optimization, medical decisions and robot coordination.
  • This reflects a strategic shift away from treating larger models and more compute as sufficient, toward methods that impose causal structure, constraints and interpretable decisions.
  • The story also reveals the role of professional rankings in China: a conference rating downgrade can alter incentives, but does not necessarily remove a venue's value for industry-relevant research.
  • The strongest claim is directional rather than quantitative: Chinese firms are seeking a theory layer for reliable AI that can survive regulated, high-consequence deployment.
Open at Leiphone Ask Claude ChatGPT

Insta360 tries to reinvent the panoramic camera after a decade

# Analysis Leiphone 18h

Chinese camera maker Insta360's X6 combines a one-inch-class sensor, three imaging chips, a gimbal mode and AI editing to make 360-degree capture less visible in daily life.

More
  • The X6 uses a custom square 1/1.1-inch sensor and claims three times the light intake of its predecessor while recording 8K at 50 frames per second.
  • Its architecture pairs a 4-nanometer Qualcomm chip with two imaging processors for noise reduction and image tuning, showing how computational photography has become central to a niche camera category.
  • The product combines a panoramic camera, a gimbal camera and first-person head tracking, reducing the need for users to choose a capture mode before recording.
  • AI director and three-dimensional time-capsule features attack the less glamorous bottleneck: reviewing and editing long panoramic footage after the shoot.
  • Insta360's history illustrates a Chinese hardware company's path from a tiny enthusiast niche to a global category leader by discovering use cases before optimizing specifications.
  • The next challenge is not sensor performance but behavior change: can software make users forget the camera and accept panoramic capture as a default memory format?
Open at Leiphone Ask Claude ChatGPT

China appears to enlarge the Type 054B frigate's missile capacity

# Analysis SCMP 10h

Satellite imagery suggests a Type 054B frigate under construction in Shanghai has a different vertical-launch layout that could increase missile capacity by about 50 percent.

More
  • The comparison is between the new vessel and the first two Type 054B frigates already in service, with imagery dated June 28.
  • A larger vertical-launch system would increase the ship's missile reserve and potentially its air-defence endurance, but the report does not establish the final loadout.
  • The story illustrates how open-source satellite imagery is being used to infer incremental changes in Chinese naval design before official specifications appear.
Open at SCMP Ask Claude ChatGPT

China's nuclear power pipeline reaches 135 gigawatts

# Signal IT Home 8h

China says 57 reactors are operating, approved or under construction, bringing the combined capacity to 135 gigawatts and keeping the country first worldwide by buildout.

More
  • Nuclear plants represent less than 2 percent of China's total installed power capacity but produce nearly 5 percent of electricity, according to the report.
  • Four coastal provinces obtain more than 15 percent of their electricity from nuclear generation, making the technology a major source of low-carbon baseload power in the east.
  • Eight new units were approved in late July, described as the first approvals of the current five-year planning period.
  • The scale reflects a state-led infrastructure model: centralized approvals, long construction cycles and concentrated coastal demand allow China to expand nuclear capacity faster than most markets.
  • The report cites a decade of top World Association of Nuclear Operators performance scores, but the safety claims are government-reported and should be read alongside independent oversight.
Open at IT Home Ask Claude ChatGPT

China's maker-tool boom faces a hard path from hobby to household

# Analysis Leiphone 15h

Chinese maker-hardware firms see a potential $36.2 billion global market by 2033, but high prices, maintenance and workflow complexity still keep 3D printers, CNC machines and UV printers niche.

More
  • The report cites a global DIY maker-tool market of $18.9 billion in 2025 and forecasts $36.2 billion in 2033, while noting that the figures come from multiple industry studies.
  • Chinese 3D-printer maker Bambu Lab is described as having shifted the category toward turnkey products and a sales target of 24 billion yuan, roughly $3.4 billion, though that target is sourced to industry contacts.
  • Desktop CNC has broader material capability than additive printing but faces dust, noise, fixturing, CAM and tool-wear problems that make the full workflow much harder for households.
  • UV printing shows the opposite failure mode: consumer demand exists, but ink and nozzle compatibility can make reliable delivery difficult.
  • The industry's shared bottleneck is component cost, with laser sources and UV printheads still dependent in part on foreign suppliers.
  • The strategic opportunity for Chinese brands is to package difficult industrial processes into appliances; the risk is that lower prices expose unresolved reliability and safety problems at much larger scale.
Open at Leiphone Ask Claude ChatGPT

Chinese game platforms move from AI creation to distribution and moderation

# Analysis 36Kr 20h

At ChinaJoy, Alibaba Cloud and TapTap said natural-language tools have made game prototyping easy; the harder problems are discovery, compliance, quality grading and platform economics.

More
  • TapTap reportedly launched more than 4,000 AI-created games in the first half of the year, then found that creators could make games faster than the platform could find players for them.
  • Alibaba Cloud frames the stack as model-as-a-service plus agentic cloud: models generate logic, while agents call tools, run sandboxes and produce playable builds.
  • The next infrastructure layer is governance, including quality evaluation, behavior auditing, identity and compliance checks for a flood of user-generated content.
  • This changes the bottleneck from capability to distribution. In China's market, where major platforms already price traffic, new creators may have no route to an audience even when production costs fall toward zero.
  • The durable product opportunity may therefore be the surrounding operating system for AI creation, not the code-generation interface that models can absorb quickly.
  • The discussion captures a broader Chinese platform lesson: lowering the cost of supply creates value only when discovery, trust and demand can scale with it.
Open at 36Kr Ask Claude ChatGPT

Chinese tech roundup: DeepSeek ships, a new AI startup seeks a $2 billion valuation

# Signal 36Kr 23h

Chinese business coverage highlights DeepSeek V4 Pro, a former Alibaba technical leader's new agent startup and investor rotation into AI hardware, while also tracking labor and retail rumors.

More
  • The roundup says former Alibaba technical leader Lin Junyang founded a Shanghai company focused on agents spanning digital and physical worlds, with an angel-round valuation reportedly near $2 billion; the figure comes from unnamed sources.
  • It reports DeepSeek V4 Pro at 3 yuan per million input tokens, roughly $0.42, and 6 yuan per million output tokens, with a million-token context window and 384,000-token maximum output.
  • The item also cites a large investor increase in Pop Mart holdings and a shift toward AI-hardware names, but those portfolio moves are investor-specific rather than sector evidence.
  • Its broader signal is personnel and capital circulation: leading Chinese model engineers are leaving large platforms, while domestic funds and strategic investors are financing new agent companies.
  • The roundup includes denials of rumored Starbucks China labor changes, illustrating how Chinese business reporting mixes verified corporate responses with fast-moving market rumor.
  • The page value lies in the ecosystem pattern, not any one unconfirmed valuation: talent, capital, model releases and hardware spending are reinforcing one another.
Open at 36Kr Ask Claude ChatGPT

Meta removes 756,000 suspected underage accounts in Australia

# Policy IT Home 7h

Chinese coverage highlights Meta's removal of 756,000 suspected under-16 accounts after Australia's social-media ban, while regulators say more than 80 percent of young users still found ways around it.

More
  • Meta says it disabled 462,000 Instagram accounts and 294,000 Facebook accounts between the eve of the ban and June, using profile and post clues such as school-year references.
  • The Australian government doubled the maximum penalty to 99 million Australian dollars, roughly $66 million, and expanded regulators' access to platform records.
  • The gap between account removals and continued use suggests that age-gating is a persistent identity and enforcement problem rather than a one-time moderation exercise.
  • For Chinese platforms, the case is a useful foreign precedent: automated inference from context can enforce age rules, but it also increases the collection and classification of sensitive personal data.
Open at IT Home Ask Claude ChatGPT

Consumer neurotechnology turns brain data into the next privacy fight

# Policy Solidot 7h

Chinese coverage warns that AI-enabled brain sensing is moving into workplaces and homes, raising demands for legal protection of mental privacy and autonomy.

More
  • The report describes consumer devices from Apple, Meta and Snap as part of a shift from laboratory neuroscience toward everyday monitoring.
  • Machine learning can now help identify emotional states in brain scans, though the supplied article does not establish the reliability or consumer readiness of those systems.
  • The policy question is broader than ordinary data protection: neural signals may reveal identity, emotion or intention in ways users cannot easily observe or revoke.
  • China is not the subject of the reported product activity, but the issue matters to Chinese regulators and companies as brain-computer interfaces move toward commercialization.
Open at Solidot Ask Claude ChatGPT

Chinese travellers mock error-filled English signs in public spaces

# Culture SCMP 15h

A mistranslated Gansu service-area sign went viral, renewing Chinese criticism of poor English signage and the gap between international tourism ambitions and local execution.

More
  • The sign rendered a hot-water room as open water rooms, part of a recurring pattern that also includes misspellings at airports and other public facilities.
  • The controversy is cultural and administrative rather than linguistic alone: public authorities are expected to project competence to domestic travelers and foreign visitors, but translation quality is often fragmented across contractors.
  • Online ridicule turns small signage failures into a test of local government professionalism and China's desire to make public infrastructure legible to outsiders.
Open at SCMP Ask Claude ChatGPT

Chinese small exporters move into AI and robotics supply chains

# Signal 36Kr 19h

A survey from cross-border payments firm XTransfer says Chinese small exporters are finding growth by supplying robot frames, AI-server cooling parts and medical packaging rather than complete products.

More
  • The company says small firms remain resilient because they can shift markets and products faster than large exporters as trade routes and demand change.
  • Its AI system TradePilot and 72 embedded agents reportedly support customer onboarding, transaction verification and anti-money-laundering review, showing AI entering compliance operations rather than only content generation.
  • XTransfer says it serves more than one million businesses and supports local collection accounts in nearly 60 countries and regions, though those are company disclosures.
  • The structural point is that China's industrial export transition is broadening beyond famous finished products: small suppliers follow anchor industries into new global niches.
  • That pattern matters for trade policy because sanctions and export controls may affect a distributed network of specialized suppliers, not just a handful of national champions.
Open at 36Kr Ask Claude ChatGPT

Google's Pixel 11 becomes a body for Gemini

# Analysis GeekPark 16h

Chinese analysis argues that Pixel 11 is designed primarily to distribute Gemini through phones, watches and future glasses, with higher prices funding more AI-oriented hardware.

More
  • The new HiLight rear ring signals Gemini's listening, thinking and answering states but has no public developer interface, making it a closed AI indicator rather than a general-purpose hardware platform.
  • Google says its Tensor G6 and Gemini Nano 4 deliver up to 3.5 times the speed and one-third the energy use for selected on-device tasks, including some offline actions.
  • The Pixel Watch 5 increases memory from 2 GB to 3 GB to support generative watch faces and Gemini, while the Pixel 11 Pro reportedly cuts base memory from 16 GB to 12 GB.
  • The analysis's comparison with China's Doubao is strategic: the handset is less a differentiated computer than a distribution channel that makes an assistant omnipresent.
  • The price increases of $100 across the range, combined with 256 GB starting storage, show Google charging for the hardware and storage needed to make AI services central to the device.
  • The unresolved business question is whether a closed Gemini-centric surface can create daily utility stronger than an open developer ecosystem built around sensors and notifications.
Open at GeekPark Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 1/20
Pekingnology 200 0/20
Ginger River Review 200 0/20
Sinocism 200 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 +41 over cap
SCMP 200 10/50 +9 over cap
Initium 200 6/15 +1 over cap
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