China Desk

China News, Summarized 28 Aug 2026 34 stories archived day

Today's brief

Removing the army's number two is now a procedural item

Zhang Youxia's seat, vacated by a standing committee

The vote took minutes. The National People's Congress Standing Committee met Friday and formally stripped Zhang Youxia and Liu Zhenli of their seats on the Central Military Commission, the body that runs the Chinese military and is chaired by Xi Jinping. Zhang was the CMC's senior vice-chairman — the highest-ranking uniformed officer in the country. Liu was chief of the Joint Staff Department, per SCMP.

Seven months, start to finish. The defence ministry announced on 24 January that the Party Central Committee had opened an investigation into both men for suspected serious violations of discipline and law. Zhang is 75, a ground-forces general, a Politburo member, one of the few serving officers with combat experience, and had been read for a decade as Xi's closest military ally.

The charge was never really corruption. A PLA Daily editorial in January said the two had trampled and undermined "the system of ultimate responsibility resting with the CMC chairman" — that is, on Xi himself. Background: He Weidong, the other vice-chairman, was expelled last October and replaced by Zhang Shengmin. Both vice-chairmen gone inside ten months.

THE TELL: there was no crisis around it. A legislative committee cleared the paperwork on an ordinary Friday, four weeks before Xi flies to Washington on 24 September. The purge that reached the top of the army has been absorbed into the calendar, which is a more remarkable claim about this system than any of the individual falls.

Go deeper on this section: ClaudeChatGPT

Beijing wires 17 provinces, Washington fences Thailand

MIIT published the node list. The ministry named 17 provinces and municipalities — Beijing, Shanghai, Hebei, Shanxi among them — as regional nodes in the national compute interconnection system. The architecture is one national service node, M regional nodes handling registration, resource pooling and scheduling, N industry nodes on top. One node per province, unified identifiers and rules.

The carriers are doing the building. China Telecom's scheduling platform already manages over 118 EFLOPS; China Mobile claims the largest single-cluster AI data centre of any operator. This is the state treating scattered provincial GPU halls as one addressable grid — the plumbing that makes "domestic compute" a thing you can rent rather than a thing you own.

Washington moved the other way. The Information reports the US is drafting a rule to close the loophole that lets Chinese firms reach AI chips through data centres in third countries like Thailand. Renting offshore was the last easy path; if the rule lands, the grid above stops being a preference and becomes the only option.

And the small countries are hedging openly. Uzbekistan asked to join Pax Silica, Washington's AI supply-chain initiative, a month after signing up for China's World Artificial Intelligence Cooperation Organization; its ambassador said Tashkent notified the State Department by diplomatic note last week, SCMP reports. Both blocs, no apology.

Go deeper on this section: ClaudeChatGPT

China's most valuable company sells DRAM

CXMT reported, and the number is absurd. First-half revenue of 150.3bn yuan (~$22.4bn), more than double everything it sold in all of 2025, and a profit of 77.6bn yuan (~$10.9bn) against a loss a year earlier — up 873.64% year on year. The results came a month after China's second-biggest Shanghai IPO, which raised 66.6bn yuan.

Read the two write-ups side by side. Bloomberg led with sales leaping tenfold and the memory maker becoming China's most valuable business. IT Home led with the guidance buried in the filing: global DRAM supply stays short through the second half, and three foreign firms still hold most of the market while domestic players have room to run. One story is an earnings beat. The other is an import-substitution scoreboard.

Then the smallest detail of the day. CXMT opened a Weibo account and the first account it followed was Xiaomi's phone division — its reported LPDDR6 partner for the Xuanjie O3, the in-house flagship SoC launched 24 August as the industry's first with LPDDR6 support. A follow button as a supply-chain announcement.

I have been under-weighting this since 14 August. Two days ago the memory squeeze took Konka, a 1980-vintage reform-era TV maker, off the Shenzhen exchange. Today the same cycle produced the most valuable listed company in the country. Same shortage, opposite ends.

Go deeper on this section: ClaudeChatGPT

Tencent's Hy4 optimised its own inference stack

The preview is open-weights. Hy4 preview: 770B total parameters, 49B active, 1M context, out today on Hugging Face and GitHub and priced at 6 yuan (~$0.85) per million input tokens, 18 (~$2.54) out, 0.3 (~$0.04) on a cache hit — per InfoQ. Two weeks free in Tencent's WorkBuddy and CodeBuddy.

The claim worth checking is the loop. Tencent says the model participated in its own development — proposing training methods, data strategies, evaluation schemes and low-level operator optimisations, running the experiments, then feeding the logs back into the next round. It analysed inference bottlenecks itself, worked on operator fusion and communication, and lifted end-to-end throughput 31.8% over baseline.

Then InfoQ gave it a procurement task and it failed on arithmetic. The agent caught genuinely obscure policy changes, then used Cursor's monthly rather than annual seat price, and separately reported a $10,800 total where the multiplication gives $10,440. A model that can tune its own kernels still cannot be trusted to cross-check a spreadsheet before signing.

Meituan drew the opposite line the same day. Chief executive Wang Xing told the earnings call that fully domestic training and inference gives long-term cost and infrastructure control, then said flatly that Meituan will not become a "token factory" — models exist to serve the core business. Its LongCat-2.0, 1.6T parameters with ~48B active, was trained end-to-end on a 50,000-card domestic cluster. Two giants, same silicon, opposite business models.

Go deeper on this section: ClaudeChatGPT

The worst film of the year is winning

"The Cow Comes" is the box-office story of the summer. A shoestring animated film five years in the making by a mother-and-son team, Sun Lifang and Xin Yumeng, it sold fewer than 300 tickets nationwide in its first ten days — about 7,000 yuan ($1,040). Clips of its astonishingly bad animation circulated, and it has now passed 40m yuan (~$5.6m), with projections above 100m.

Screenings have become events. Audiences recite the dialogue back at the screen and film themselves doing it; theatres caught without official posters drew their own by hand, which then became their own genre of online joke. China Digital Times has the translations.

The backlash is the interesting part. CCTV-affiliated outlets criticised it, The Beijing News called it nonsensical, Dalian officials reportedly worried about the city's reputation, and there are rumours the film could simply be pulled. One Xiaohongshu user put the asymmetry cleanly: "'Ne Zha' was called 'a blockbuster,' but 'The Cow Comes' is being called 'an anomaly.'"

The same afternoon, the industrial hit crossed 1.8bn yuan. "Welcome to Dragon Restaurant" passed 1.8bn yuan (~$254m) at 9:57pm. Nobody is arguing about that one. The film people are arguing about cost nothing to make and was funny by accident, which is exactly why it cannot be permitted to be the story of the summer.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The Alliance sentencing landed where I said it would. On 22 August I flagged 28 August. Today: Lee Cheuk-yan and Ho Chun-yan, former leaders of the group behind Hong Kong's annual Tiananmen vigil, argued in mitigation for lighter terms; Chow Hang-tung, the barrister who ran her own defence, refused to plead inside a criminal framework at all, saying she should not serve a single day.
  • Gyirong now has video of the cause. Drone footage posted to RedNote and verified by the BBC shows the Langtang Lirung glacier collapsing at about 23,600 feet. Nepal reports 547 dead and 575 foreign tourists missing; Tibet still reports 5 dead and 558 missing, 260 of them foreign. The barrier lake remains a live breach risk.
  • A Beijing satellite firm redirected six radar birds within hours, imaging the collapse site through cloud — the same information vacuum that produced a wave of posts blaming Chinese-built infrastructure for a glacier failure on the Nepali side.
  • MiniMax's $800m ARR is now on the record. I cited it Wednesday off the results call; the half-year report confirms August ARR past $800m, B-side revenue up 703% to 80% of the mix, and July token consumption 20x January. Next up: M3 Pro at roughly 3T parameters. Zhidx has the framing — minimise inference cost to maximise intelligence.
  • AgiBot's actual boss surfaced on a US list. Time's 100 AI names 11 Chinese figures, including Deng Taihua — 20-plus years at Huawei running wireless then the computing line, who recruited the Bilibili-famous engineer Peng Zhihui and then spent two years letting him take the cameras. AgiBot shipped 5,100 robots last year.

Go deeper on this section: ClaudeChatGPT

The river · 34

China removes two former generals from its top military command

# SCMP

Beijing formally stripped Zhang Youxia and Liu Zhenli of Central Military Commission positions after months of investigation, a major purge at the top of the PLA.

More
  • Zhang was the highest-ranking uniformed officer in China's military, making his removal more consequential than a routine personnel change.
  • The two men had reportedly been under investigation for discipline violations since January, and the formal action came through the National People's Congress Standing Committee.
  • The move demonstrates that the party-state can use legislative institutions to formalize a military personnel decision after an internal investigation.
  • A purge at this level may affect command continuity, procurement oversight, promotion networks, and perceptions of loyalty inside the armed forces.
  • It also raises questions about the status of the wider military leadership and whether additional removals or reshuffles will follow.
  • The supplied report does not state the precise allegations or identify successors, so the cause and operational impact remain unclear.
  • For external observers, the key signal is not only who was removed but how much authority remains concentrated around the top leadership during a period of military modernization.
Open at SCMP Ask Claude ChatGPT

Xi-Trump summit plans point to another trade truce extension

# SCMP

Chinese reporting says Xi Jinping may bring a business delegation to Washington for a September summit with Donald Trump, where both sides may extend their trade-war truce.

More
  • The report cites four people familiar with negotiations and says the delegation is still under discussion, so the trip and agreement are not final.
  • A business delegation would signal that Beijing and Washington are trying to pair leader-level diplomacy with corporate and commercial reassurance.
  • The likely truce extension would preserve a managed pause rather than resolve the structural disputes over technology, market access, and industrial policy.
  • The comparison with Trump's May China visit, when US executives such as Tim Cook and Elon Musk traveled with him, underscores the role of business elites as diplomatic instruments.
  • For technology companies, the immediate implication is continued uncertainty rather than normalization: export controls, investment review, and supply-chain decisions remain hostage to the next political bargain.
Open at SCMP Ask Claude ChatGPT

China's top legislature passes five laws and removes senior officials

# Jiemian

China's National People's Congress Standing Committee passed five laws and removed senior civilian and military officials, including two Central Military Commission members.

More
  • The session passed laws on medical insurance, farmland protection, agriculture, and national defense mobilization, plus amendments to the Lawyers Law.
  • It removed the civil affairs minister Lu Zhiyuan and appointed Li Changguan, alongside several changes at the courts, state supervision commission, and legislature.
  • More consequentially, Zhang Youxia was removed as vice chairman of the Central Military Commission and Liu Zhenli as a commission member.
  • The announcement provides no explanation for the removals, so the personnel changes are more important than the routine legislative package but remain difficult to interpret.
  • The pattern reinforces how formal legislative sessions function as a vehicle for announcing elite personnel decisions as well as laws.
  • For outside observers, the absence of stated reasons is itself part of the signal: personnel changes are public, while the underlying political process remains opaque.
Open at Jiemian Ask Claude ChatGPT

China's auto regulator draws a quality line under 18-month car development

# TMTPost

Four ministries launched a year-long inspection campaign covering reliability, suppliers, driver assistance and AI safety, explicitly examining whether development cycles are too short.

More
  • The campaign reaches behind homologation results into design life, research timelines, supplier controls and validation of new materials and software.
  • China's electric-car developers have compressed new-model cycles from roughly 40-50 months in traditional programs to 15-18 months in some cases.
  • The regulator is not imposing a minimum development period; it is distinguishing platform reuse and simulation from skipped durability, supplier or extreme-condition testing.
  • That matters because software defects can be patched over the air, while fatigue, suspension and structural failures cannot be repaired after sale.
  • Unannounced inspections could make previously invisible engineering compromises auditable, shifting speed from a marketing metric toward a liability question.
  • The policy will matter most if it changes launch incentives and supplier documentation rather than merely adding another inspection checklist.
Open at TMTPost Ask Claude ChatGPT

Anthropic's MHS turns lab devices into agent-callable tools

# AI QbitAI

Anthropic says its Model Hardware Standard lets agents discover and control diverse laboratory devices, from microscopes to robotic arms, without task-specific training or bespoke integration.

Also: Zhidx, Leiphone

More
  • The standard exposes device capabilities, state and safety limits through machine-readable descriptions and common driver operations.
  • In a demonstration, Claude reportedly measured and calibrated a low-cost robotic arm before controlling it, while other tests automated microscopy and scientific workflows.
  • The concept extends MCP from software tools into a physical world where bad commands can damage equipment or invalidate experiments.
  • MHS remains a limited preview; open-source release, safety validation and real vendor support will determine whether it becomes infrastructure or a compelling demo.
Open at QbitAI Ask Claude ChatGPT

China's Yunzhisheng turns enterprise agents into a growing business

# Signal QbitAI

Yunzhisheng says first-half agent revenue reached 478 million yuan, roughly $67M, while token revenue rose 760%, showing a Chinese enterprise-AI model built around deployment and repeat purchases.

More
  • Total first-half revenue rose 38.7% to 562 million yuan, roughly $79M, with agent work accounting for 85.1% of the total.
  • The company says it serves more than 470 medical institutions, has built 524,400 enterprise profiles in Xiamen, and operates across healthcare, insurance, transport, and manufacturing.
  • Token revenue was still small at about 30 million yuan, roughly $4.2M, but its 60%-plus gross margin and rapid growth suggest customers are beginning to pay directly for model access.
  • More than 60% of revenue reportedly comes from repeat purchases, which is a better signal of workflow value than one-off deployment contracts.
  • The commercial model resembles an industrial AI integrator: base models provide capability, while applications, data, and system integration create switching costs.
  • The company remains loss-making, so the strategic question is whether gross-margin growth can outrun delivery and research costs.
  • For China's AI market, the case supports a less glamorous thesis: durable revenue may come from embedding agents into regulated operations rather than from consumer novelty.
Open at QbitAI Ask Claude ChatGPT

Zhipu launches a low-cost multimodal model on domestic compute

# AI QbitAI

Zhipu open-sourced GLM-5.3-Flash, a 320B model with 18B active parameters, while SenseTime says its domestic-chip service handled 62T test tokens and reached NVIDIA-like efficiency.

More
  • The model is priced at 6 yuan per million input tokens (about $0.85), 18 yuan per million output tokens (about $2.54) and 0.3 yuan per million cached input tokens (about $0.04).
  • SenseTime reports threefold end-to-end performance improvement over its initial domestic-chip baseline and claims domestic hardware efficiency comparable to NVIDIA GPUs; these are sponsored claims.
  • The implementation uses heterogeneous inference so chips with different compute and bandwidth profiles can handle different phases instead of forcing a single architecture to do everything.
  • SenseTime says its token factory served 2.42 trillion tokens per day in July and may reach 10 trillion by year-end, figures that require independent verification.
  • The concrete signal is that Chinese model companies are testing domestic accelerators under large real traffic, moving the debate from isolated benchmarks toward service economics.
Open at QbitAI Ask Claude ChatGPT

Hermes tests agents that carry a user's browser identity

# AI Zhidx

A Chinese review finds Hermes can copy a Chrome profile into an isolated browser, but encryption, cookies, and background processes make the feature both fragile and high risk.

More
  • The feature copies login state rather than taking over the active browser, allowing an agent to access sites where the user is already authenticated.
  • On macOS, copied credential files remained encrypted and could not always be used; cookie replacement and browser-process conflicts caused additional failures.
  • The security boundary is serious: an agent acting with a user's identity can create consequences that ordinary API permissions and visible browser sessions would make easier to audit.
  • Hermes, Claude, and Codex all illustrate the same product direction, moving from tool invocation toward delegated identity and computer use.
  • The review found long runtimes and bugs, so convenience is not yet a reason to accept invisible background actions.
  • For enterprise deployment, the unresolved questions are consent, action logs, revocation, least privilege, and liability when a session acts incorrectly.
Open at Zhidx Ask Claude ChatGPT

MiniMax argues that cheaper inference can buy more intelligence

# Signal Zhidx

MiniMax says first-half revenue grew 283.1%, ARR exceeded $800M in August, and 80% of revenue came from business customers, tying its growth strategy to lower unit inference cost.

Also: New Wisdom

More
  • The company's reported token consumption in July was 20 times January's level, suggesting that usage rather than only subscription count is driving expansion.
  • The core argument is that inference is also the raw material for reinforcement learning, synthetic data, evaluation, and agent interaction, so efficiency expands both product margins and research capacity.
  • This rejects the simple tradeoff between cheap small models and expensive frontier models by making cost reduction part of capability development.
  • The figures are company disclosures and the report does not provide audited customer counts, gross-margin detail, or independent benchmark methodology.
  • For Chinese model vendors, the commercial path may favor high-volume APIs and enterprise workflows if they can deliver adequate intelligence at much lower cost.
  • The main risk is that frontier workloads continue to demand large models and erase efficiency gains through longer agent trajectories.
Open at Zhidx Ask Claude ChatGPT

OpenAI's Jalapeno makes token efficiency a chip battleground

# AI Leiphone

OpenAI's reported Jalapeno inference chip targets tokens per watt and decode latency, reflecting a hardware split between compute-heavy prefill and memory-bound generation.

More
  • OpenAI reports 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower end-to-end latency across several large models, with higher gains in interactive ranges.
  • The architectural focus is the decode phase, where weights and KV cache must move repeatedly while parallelism is low; peak matrix throughput alone cannot solve that bottleneck.
  • The comparison with NVIDIA Rubin is disputed because software, load shape and system configuration are not aligned, and Jalapeno is not yet a production-scale deployment.
  • NVIDIA's use of a Groq inference processor and Google's separation of training and inference TPU paths suggest a wider industry move toward workload-specific systems.
  • For AI infrastructure buyers, tokens per watt, tail latency and cache locality may become more useful procurement metrics than nominal FLOPS.
Open at Leiphone Ask Claude ChatGPT

Tencent's Hy4 preview builds a 3D game from one prompt

# AI Zhidx

Tencent's Hy4 preview uses 770B total parameters, 49B active and a 1M-token context to handle long coding and game-building tasks, though its impressive demos remain vendor tests.

Also: InfoQ China

More
  • The model reportedly created a playable low-poly open-world game over an hour, fixing JavaScript and rendering problems during its own validation loop.
  • Its API price is 6 yuan per million input tokens (about $0.85), 18 yuan per million output tokens (about $2.54) and 0.3 yuan per million cached input tokens (about $0.04).
  • The publisher says Hy4 improved long-context decomposition, tool use and self-correction over Hy3, with the model open-sourced for local and hosted use.
  • A 1M context and a polished demo do not prove reliable software maintenance; the model still depends on a harness, permissions and human acceptance testing.
  • This is the same Hy4 release covered elsewhere, so it belongs to the same event cluster rather than being treated as a separate model launch.
Open at Zhidx Ask Claude ChatGPT

Hong Kong activists seek leniency while Chow rejects the case's frame

# Policy Initium

Hong Kong Alliance defendants asked for lighter sentences, while Chow Hang-tung argued that she should serve no prison term, exposing different strategies toward political prosecution.

More
  • The case concerns the former Hong Kong Alliance, whose leaders were convicted under the national-security-era sedition framework described in the supplied report.
  • The contrast between mitigation and outright rejection is politically meaningful: one strategy seeks reduced punishment within the court's frame, the other contests the frame itself.
  • The proceeding shows how legal risk has reshaped civil-society organizations and the language available to defendants.
  • The supplied item is a digest excerpt and does not provide the final sentences or the full court reasoning.
Open at Initium Ask Claude ChatGPT

Hong Kong bookstore arrests reopen a painful Taiwan debate

# Culture Initium

Chinese-language independent bookstores in Taiwan rallied around the phrase I am a bookstore worker after Hong Kong arrests, but the gesture also exposed unresolved tensions between Hong Kong and Taiwan.

More
  • Hong Kong police reportedly searched two independent bookstores and arrested five people under the sedition law, turning a staff shirt into a widely shared symbol.
  • The phrase moved from Hong Kong into Taiwan bookstores, concerts, and political posts, where some saw solidarity and others saw political consumption.
  • The dispute shows how shared memories of the 2019 protests have diverged as Hong Kong's legal and political environment changed.
  • Bookstores matter here as civic infrastructure: restrictions on them affect not only commerce but the circulation of history, identity, and dissent.
  • The podcast framing is interpretive rather than a court report, but it captures how cultural gestures become proxies for cross-strait and intercity trust.
  • The continuing question is whether Taiwan can express solidarity without flattening Hong Kong's distinct risks and agency.
Open at Initium Ask Claude ChatGPT

A cross-model flaw exposed hidden AI reasoning blocks

# Policy Leiphone

Researchers reportedly recovered readable hidden reasoning by moving signed blocks between models, suggesting some APIs verify integrity without binding state tightly enough to model, account or session.

More
  • The design returns encrypted reasoning to the client so later calls can resume without storing all state server-side, but encryption alone proves authenticity rather than current authorization.
  • The reported attack gives a block produced by a stronger model to a more permissive model from the same provider, which may then transcribe content it was never meant to expose.
  • The issue is a context-binding failure, not a cryptographic break: the system accepts a valid block in the wrong model or session.
  • Cross-session resumption and model switching are legitimate product requirements, so the fix must distinguish permitted portability from unrestricted replay.
  • The incident matters for agent APIs because opaque reasoning, tool state and user identity are increasingly passed through clients and orchestration layers.
Open at Leiphone Ask Claude ChatGPT

Chinese analysts question whether vision-based touch is a real moat

# AI QbitAI

A Chinese industry analysis argues that camera-based tactile sensors rely on mature components and borrowed vision algorithms, making them easy to copy and difficult to scale across robot bodies.

More
  • The critique targets vision-based tactile sensors built from a camera, light source, elastic material, and markers, rather than tactile sensing as a whole.
  • The article says most systems remain thicker than 10mm and depend on externally supplied CMOS sensors, limiting integration into compact dexterous hands.
  • Because their algorithms often reuse computer-vision methods, many vendors may struggle to differentiate on accuracy, durability, or long-term calibration.
  • The strongest industrial objection is deployment geometry: rigid optical modules fit fingertips but do not naturally form flexible, curved electronic skin across a robot.
  • The analysis is deliberately adversarial and does not provide comparative field-failure data, cost curves, or evidence that all vision-based approaches share the same limits.
  • For investors and builders, the useful test is not demo resolution but lifetime stability, repairability, supply-chain control, and value in a complete manipulation loop.
Open at QbitAI Ask Claude ChatGPT

Alibaba's ABot-Recon rebuilds long 3D scenes from 12 frames

# AI QbitAI

Alibaba's Amap released a streaming 3D-reconstruction model that uses a 12-frame local window to build scenes over thousands of frames, cutting memory growth without long-range anchors.

More
  • ABot-Recon predicts local point clouds and relative camera pose, then composes them online into a global trajectory rather than storing the entire sequence in memory.
  • The source reports 24.45 frames per second on KITTI, 6.71GB peak memory, and a 40.6 percent lower trajectory error than a representative method on a long Oxford sequence.
  • It uses only monocular RGB video and does not require depth sensors or known camera parameters, lowering deployment requirements for private spaces such as warehouses and malls.
  • The counterintuitive design is that bounded local context can produce more stable long-run computation than a large memory bank, provided drift correction is strong enough.
  • The remaining risk is domain transfer: camera motion, scene changes, dynamic objects, and poor texture can still break the local-pose assumptions.
  • Alibaba has released code and weights, making this one of the more directly testable Chinese physical-AI releases in the batch.
Open at QbitAI Ask Claude ChatGPT

Cursor's Origin redesigns code hosting for agent-scale traffic

# AI Leiphone

Cursor's Origin beta combines repositories, pull requests, checks, reviews and automation around continuously running agents, challenging GitHub's human-paced control plane.

More
  • A demo showed 22.6 commits per second in one repository, which the article correctly labels as a demo rather than a production benchmark.
  • Agent traffic changes the bottleneck from Git objects to APIs, webhooks, CI queues, branch protection, indexing and mergeability state.
  • A study cited in the article found overlapping agent pull requests in 40.2 percent of sampled repositories and a 41.7 percent text-conflict rate among replayed cross-agent changes.
  • Origin's proposed small, stacked changes act as checkpoints for local validation and retry, reducing the cost of recovering from one failed step.
  • The strategic threat to GitHub is not a better code editor but a repository service designed around machine-speed concurrency and state propagation.
  • The missing proof is whether Cursor can operate that control plane reliably at scale without recreating the same human-oriented bottlenecks.
Open at Leiphone Ask Claude ChatGPT

US and China turn war messaging into meme warfare

# Culture ChinaTalk

A China-focused analysis argues that both countries are packaging military power through cute filters, video-game aesthetics, and short-form propaganda that blurs entertainment and threat.

More
  • The article contrasts Chinese state-media animal transformations into missiles, drones, and robots with US military highlight reels using action-movie editing.
  • The medium matters because memes reduce the emotional distance to violence while making strategic signaling more shareable and harder to distinguish from ordinary entertainment.
  • Chinese military messaging is not simply decorative: the animal-to-weapon transformations imply coordinated force across air, sea, ground, and unmanned domains.
  • The US-China comparison is analytical and polemical, and the supplied examples do not establish how many people believed or acted on the content.
  • The social risk is normalization: repeated aesthetic treatment of war can make escalation feel like a game and obscure civilian consequences.
  • For technology executives, the relevant infrastructure is recommendation and generative-media systems that can scale this rhetoric across borders.
Open at ChinaTalk Ask Claude ChatGPT

Alibaba opens Qoder's coding agent to non-programmers

# AI Zhidx

Alibaba's redesigned Qoder hides code behind natural-language collaboration while retaining model choice, tools and enterprise connectors for professional developers.

More
  • The platform separates programming and general modes, with an automatic dispatch engine balancing model quality, speed and cost for users who do not want to select a model.
  • Alibaba says Qoder has more than six million users, over 100,000 enterprise users, 40-plus connectors, 70-plus plugins and a 20,000-plus skill library; these are company figures.
  • Publisher tests included packaging a desktop image-processing tool and generating an iOS health app, with browser screenshots and functional checks feeding an iterative loop.
  • The strategic move is distribution: coding agents become a general execution layer for designers, operators and analysts rather than a specialized editor for engineers.
  • The unresolved risk is that hiding implementation raises the need for permissions, provenance, testing and maintenance interfaces for nontechnical owners.
Open at Zhidx Ask Claude ChatGPT

The next cancer-vaccine breakthrough may cost nearly $700,000

# Analysis QbitAI

Chinese coverage warns that a personalized mRNA melanoma vaccine could cost $475,000 before companion therapy, turning an AI-assisted clinical milestone into an affordability debate.

More
  • The $475,000 figure is an investment-bank forecast, not an announced price, and the vaccine has not yet been approved.
  • The treatment is individualized: tumor sequencing, neoantigen selection, mRNA design, manufacturing, and quality control are repeated for each patient.
  • Adding up to nine cycles of pembrolizumab could bring the modeled drug cost close to $696,000, roughly $98,000, but actual coverage and pricing remain unknown.
  • Chinese analysis usefully separates AI's contribution from the rest of the platform, including clinical trials, lipid delivery, manufacturing, and regulatory infrastructure.
  • The social question is whether personalized therapies become accessible through insurance and manufacturing scale or remain premium products for wealthy patients.
  • This is a foreign medical story, but the Chinese framing adds a substantive affordability and health-system perspective.
Open at QbitAI Ask Claude ChatGPT

A Chinese robot learns to open doors through one video and simulation

# AI Leiphone

Researchers from Niewa Robotics, Shanghai Jiao Tong University, and Shandong University report a single-video digital-twin pipeline that reached 96.57% success on five real doors and 80.95% zero-shot success on similar unseen doors.

More
  • The system reconstructs a door's geometry, handle, hinge, texture, and collision model from an RGB video, then generates and filters trajectories in simulation.
  • Training randomizes friction, damping, camera calibration, initial pose, and depth noise so the policy must survive physical variation before reaching the robot.
  • The robot executes locally using front and wrist depth views plus state, coordinating base, arm, and gripper through approach, handle rotation, pushing, and traversal.
  • The 96.57% figure covers 169 successes in 175 tests on five target doors; the 80.95% figure covers structurally similar doors unseen during training.
  • Those results demonstrate a promising sim-to-real path but not general door opening: pulling doors, unfamiliar handles, and broader buildings remain untested.
  • The industrial metric to watch is adaptation cost per new facility, including video capture, asset correction, real-world trials, and human intervention.
  • The work is strategically relevant because it shifts scene onboarding from manual modeling toward reusable digital twins and generated experience.
Open at Leiphone Ask Claude ChatGPT

Tencent open-sources Hy4 and targets the first tier of open models

# AI Leiphone

Tencent's Hy4 Preview combines 770B parameters, 49B active parameters, and 1M context with claimed gains on real production tasks and a 31.8 percent inference-throughput improvement.

Also: Techmeme

More
  • Tencent's blind test of 203 engineering tasks with 163 internal experts gave Hy4 a slightly higher average score than GLM 5.3 and Kimi K3, but the test was internal.
  • The model is optimized for coding, office analysis, game development, and scientific work, and is distributed through Tencent products, TokenHub, and OpenRouter.
  • The company says Hy4 participated in its own data, evaluation, training, and operator optimization loop, an early form of recursive self-improvement rather than autonomous continual learning after release.
  • API pricing is 6 yuan per million input tokens, roughly $0.85, and 18 yuan per million output tokens, roughly $2.54; cached input is 0.3 yuan, about $0.04.
  • Chinese coverage presents the release as a product-and-infrastructure co-design story, while the US-facing item reduces it to Tencent's unverified internal superiority claim.
Open at Leiphone Ask Claude ChatGPT

StreamPI gives Chinese robot models a continuous sense of time

# AI Leiphone

Daxiao Robotics and the University of Hong Kong report StreamPI, a streaming temporal architecture that improves long-horizon robot tasks without adding model parameters.

More
  • Instead of repeatedly feeding a window of past frames, StreamPI binds the task instruction to each observation and stores prior representations in a key-value cache.
  • Random time-interval training helps the model cope with real sensors and actuators that do not operate at a fixed cadence.
  • The reported results improve LIBERO long-task success from 92.4% to 95.0% and raise CALVIN average continuous-task length from 4.313 to 4.547.
  • The key conceptual change is from recognizing a frame to modeling state evolution: what moved, what was occluded, and whether the task is progressing.
  • The method is lighter than adding a separate video encoder or memory module, but benchmark gains near saturation do not yet establish broad household or industrial robustness.
  • The real test is whether streaming memory reduces recovery failures when objects move unpredictably and observation timing is irregular.
Open at Leiphone Ask Claude ChatGPT

WALL-SS tries to make robot world models causal

# AI GeekPark

Chinese startup Zizaitian's WALL-SS predicts future robot states from actions through a next-scale autoregressive design, reporting better 60-second rollouts and a 0.926 sim-to-real success correlation.

More
  • Rather than generating visually plausible futures with diffusion alone, the model organizes observations and actions as an observation-action-new-observation sequence.
  • A coarse-to-fine hierarchy is intended to preserve object and robot state before adding motion and contact detail, reducing the magnetic-grab shortcut where an unclosed gripper still appears successful.
  • The supplied tests report an action-following score of 0.29 versus 0.044 for Cosmos3-Nano, with 60-second rollouts retaining more stable trajectories.
  • In 600 paired virtual and real experiments, the company says simulated and real task success correlated at 0.926 and strategy ranking was 89% accurate.
  • These are the right validation questions for a robot world model: does it obey actions, remain stable over time, and predict which policy is better rather than merely look realistic?
  • The remaining risk is distribution shift in materials, contacts, occlusions, and robot bodies; a strong benchmark correlation does not yet replace physical trials.
  • If the results generalize, world models could become an evaluation and planning layer that reduces expensive real-robot experimentation.
Open at GeekPark Ask Claude ChatGPT

Meta designs a 30B local agent for 24GB GPUs

# AI Leiphone

Meta's Muse Glimmer combines grouped-query attention, local-global attention, quantization and DFlash to run a long-context multimodal agent within consumer-class memory.

More
  • The model has 52 dense Transformer layers, 32 query heads but two KV heads, reducing cache size; only global-attention layers retain the full 128K context.
  • The publisher estimates a full-context KV cache near 1.7 GiB under the mixed design, versus more than 20 GiB with conventional per-head KV storage; this is a theoretical estimate.
  • A 17GB 4-bit weight package plus vision and acceleration components targets 24GB cards, while a larger quantization offers lower reported accuracy loss.
  • The design shows that local deployment is a systems problem: attention topology, cache policy, weight precision and decode acceleration must fit together.
  • A 128K input is not the same as durable agent memory; deciding what to retain and discard remains a runtime responsibility.
  • Open Apache 2.0 licensing and local execution could make this architecture more useful to developers who cannot send screen and tool state to the cloud.
Open at Leiphone Ask Claude ChatGPT

SenseTime says video generation now runs on domestic chips

# AI QbitAI

SenseTime and HiDream report moving a video-generation workload to Chinese accelerators through unified hardware abstraction, multi-card parallelism and deployed engineering support.

More
  • The migration used step and classifier-free-guidance distillation plus sequence and tensor parallelism, with the partners reporting a 93 percent multi-card speedup.
  • A hardware abstraction layer with registries and operator plugins is said to support migration across more than ten heterogeneous Chinese chip platforms.
  • The work also addressed output drift between single-card and multi-card inference, including face-feature consistency and frame-count handling.
  • This is a vendor-sponsored case study, so the meaningful proof is that the workload reached production traffic, not the promotional speed figure alone.
  • The pattern matters for China's software stack: portability depends on kernels, communication and toolchains as much as on model weights.
Open at QbitAI Ask Claude ChatGPT

Chinese radar satellites map the Nepal glacier disaster within hours

# Signal SCMP

Chinese satellite operators redirected six radar satellites to the disaster zone, while Chinese coverage stresses rapid commercial response and all-weather terrain mapping.

Also: Hacker News

More
  • Radar can observe through cloud and haze, making it useful for rapid mapping in the Himalayan plateau where optical imagery may be blocked.
  • The response shows commercial Chinese space companies functioning as emergency-observation infrastructure rather than waiting for a state-only tasking chain.
  • The source does not identify the company, image resolution or whether the data was delivered to rescuers, so operational impact remains unclear.
  • The event is part of the wider China-Nepal flood and should not be read as evidence that satellite imagery alone solves mountain early warning.
Open at SCMP Ask Claude ChatGPT

A terrible Chinese animated film becomes a $6 million meme

# Culture China Digital Times

The low-budget animated film The Cow Comes grew from fewer than 300 tickets to more than 40 million yuan, about $5.6 million, as audiences bought tickets to laugh at its awkwardness.

More
  • The film was made over five years by a mother-and-son team and initially earned about 7,000 yuan, roughly $1,000, before clips of its crude animation went viral.
  • Screenings became participatory performances, with audiences reciting dialogue, talking back to the screen and sharing hand-drawn theater posters online.
  • The backlash, including criticism from state-linked outlets and reports of local intervention, shows the limits of official cultural gatekeeping when irony becomes the product.
  • The hit is less a conventional quality success than a consumer referendum on viral discovery, collective mockery and the pleasure of reclaiming a bad object.
Open at China Digital Times Ask Claude ChatGPT

Claude Code sessions can now message one another

# AI Leiphone

Anthropic is testing cross-session messaging that lets separate Claude Code agents exchange text through local registration and sockets without merging their full contexts.

More
  • The design keeps sessions isolated and sends only task results or dependency information, avoiding a single shared context that would mix unrelated code and logs.
  • Agents discover one another through names, working directories and inbox sockets; local communication stays on the machine unless a remote session is involved.
  • Messages wait between tool calls rather than interrupting an active operation, which preserves execution semantics but introduces delayed coordination.
  • The architecture resembles service-to-service messaging more than context sharing, making explicit interfaces and ownership important for multi-agent codebases.
  • Security depends on filesystem visibility, operating-system permissions and container boundaries, so deployment topology becomes part of the agent protocol.
Open at Leiphone Ask Claude ChatGPT

The China-Nepal flood raises questions for a proposed rail link

# Policy SCMP

Chinese analysis argues that the border flood should prompt joint early-warning systems before a proposed China-Nepal railway advances through the same exposed mountain corridor.

More
  • The Gyirong crossing handles about one-third of China's trade with Nepal and was cut off by the mudslide, making the disaster a direct test of the corridor's resilience.
  • The proposed railway would face hazards that roads and ports already experience: flash floods, debris flows, limited access and difficult monitoring.
  • The article frames early warning as a condition of infrastructure legitimacy, not merely an emergency-response add-on.
  • The source proposes a policy direction but gives no engineering design, financing decision or agreed bilateral mechanism.
Open at SCMP Ask Claude ChatGPT

US sanctions a Hong Kong firm over alleged Iran money laundering

# Policy SCMP

Chinese coverage reports new US sanctions on a Hong Kong company and links them to a wider campaign against Chinese entities, while Beijing threatens retaliation before a possible Xi visit.

More
  • The US Treasury alleges the company helped launder funds for an Iranian exchange house, making Hong Kong-based financial infrastructure part of the enforcement target.
  • The story shows how sanctions pressure can reach Chinese jurisdictions through financial intermediaries rather than only through mainland technology companies.
  • Chinese retaliation language gives the episode a diplomatic dimension: Iran-related enforcement is being folded into broader US-China bargaining.
  • The supplied excerpt does not identify the firm's services, ownership structure, or Beijing's specific planned measures, so the legal basis and economic effect remain unclear.
  • For companies operating through Hong Kong, the practical risk is exposure to secondary-sanctions networks even when the underlying business is not directly Iranian.
Open at SCMP Ask Claude ChatGPT

Agent workloads reopen the market for Chinese server CPUs

# AI Leiphone

Chinese chip investors see a five-year CPU window because agents spend much of their time orchestrating tools, querying databases, and moving data outside the GPU.

More
  • One server vendor estimates that more than 80% of a representative agent workflow's load runs on CPUs, though the figure is an industry observation rather than a standardized benchmark.
  • The article says CPU-to-GPU ratios in AI centers could move from roughly 1:8 toward 1:4 or even 1:1 as multi-step inference expands.
  • Server CPU prices rose nearly 30% across the cited period, but Chinese experts disagree over whether the increase reflects durable value or temporary supply imbalance.
  • Established Arm products from Huawei and Hygon have the advantage of volume, software compatibility, and procurement relationships, while startups must reach production quickly.
  • RISC-V attracts investment for autonomy and licensing certainty but still lacks the ecosystem and customer orders needed for broad server deployment.
  • The real contest is not CPU versus GPU; it is which vendors can make heterogeneous systems easy to deploy across domestic accelerators, databases, and agent runtimes.
  • Agent traffic could create durable CPU demand, but the long-term winner will be determined by total system cost, memory, networking, and software support.
Open at Leiphone Ask Claude ChatGPT

Cursor's MoK targets the communication tax in MoE training

# AI Leiphone

Cursor open-sourced Mixture-of-Kittens, a GPU kernel that combines token routing, cross-GPU communication and expert computation to reduce MoE scheduling overhead.

More
  • Mixture-of-experts systems pay for dispatch, layout, synchronization and combine operations in addition to the matrix multiplication itself, even on high-bandwidth NVLink systems.
  • MoK changes dispatch from push to pull so destination GPUs organize incoming tokens locally, trading some bytes for less coordination among senders.
  • The kernel also overlaps communication and computation more deliberately, addressing small uneven expert batches that leave tensor cores or network links waiting.
  • The lesson is architectural rather than vendor-specific: after hardware bandwidth improves, software orchestration can become the limiting factor.
  • The reported microbenchmarks and kernel results do not yet establish production training gains across model sizes, network topologies or failure modes.
Open at Leiphone Ask Claude ChatGPT

Qwen cuts office-agent token use by 75%

# AI InfoQ China

Alibaba's Qwen3.8-Flash and a new standard mode for Qwen Office reportedly halve task time and cut average token use 75%, making agent economics a product-design problem.

More
  • The model is a multimodal mixture-of-experts system with 125B total parameters, 6B active per token, and an additional 51B n-gram embedding component.
  • Alibaba reports input pricing of 0.8 yuan per million tokens and output pricing of 2.7 yuan, roughly $0.11 and $0.38 respectively.
  • The more important optimization is outside the model: context management, tool orchestration, and an agent harness reduce wasted calls and let a cheaper mode handle roughly 95% of routine tasks.
  • A comparison of standard and advanced modes found similar completion for a basic web-page task, though the advanced mode handled details more fully.
  • This suggests agent products may win by routing everyday work to an adequate model rather than maximizing intelligence on every request.
  • The claim remains company- and publication-tested rather than independently benchmarked across a broad task distribution.
  • For enterprise buyers, the missing metrics are failure rates, review time, and cost after human correction, not just token consumption.
Open at InfoQ China Ask Claude ChatGPT