China Desk

China News, Summarized 11 Nov 2026 34 stories archived day

Today's brief

Nvidia's scandal is now an export-control case, and Chinese outlets are borrowing American voices to say so

The affair turns into an enforcement case

A prosecutor now runs the inquiry. Nvidia's board hired Joon H. Kim, a former acting US attorney in Manhattan, and his firm Wachtell Lipton to investigate whether chief executive Jensen Huang was targeted by foreign intelligence. A shareholder class action followed minutes after Monday's close. The stock fell over 5% Tuesday and over 9% across two days, erasing more than $420 billion.

The Commerce stall is the deeper problem. The New York Times reported that Commerce's export-enforcement bureau stalled its probe of Megaspeed. This is the Singapore data-centre firm co-founded by Alice Huang, whom US officials believe worked for Chinese intelligence. The stall followed an August 2025 email from Nvidia's general counsel. Justice then asked Commerce to pause. Nvidia says it routinely consults officials; Commerce says it followed procedure.

Chinese outlets chose the enforcement frame. TMTPost, a Chinese tech outlet, translates a CNBC interview with Dmitri Alperovitch, the CrowdStrike co-founder who chairs the think tank Silverado. His verdict: Huang cannot survive as CEO, and export-control exposure, not the affair, could put him in court. TMTPost adds that none of it is proven. Alice Huang appears only as someone the US suspects of intelligence ties. The Ministry of State Security never appears.

The market story is succession. Solidot, a Chinese tech-news site, leads on key-person risk instead. Nvidia has no obvious successor, and analysts see a 25–30% drop if Huang goes.

The counter-reading. Desmond Shum, the Chinese businessman turned author of Red Roulette, argues that Alice Huang is likelier a commercial "white glove" fixer than a spy. SCMP's explainer relays Chinese experts warning that hawks will turn the case into another "spy balloon" saga. My read: the enforcement frame is right on substance and convenient for Beijing. It puts Washington's rule-keepers, not Beijing's agents, in the dock.

WHAT WE'RE WATCHING: Nvidia's earnings on 17 November, then the closed-door House briefing the next day.

Go deeper on this section: ClaudeChatGPT

Beijing keeps 100 ships quiet before Shenzhen

The drill nobody named. Taiwan's security chief Joseph Wu says more than 100 PLA Navy and coast guard ships are at sea. Defence minister Wellington Koo confirmed "area denial" drills around the Liaoning carrier group. Beijing has declared nothing. Its defence ministry called the carrier's training routine last month.

A contested reading, and a silence. Two US think tanks, the Institute for the Study of War and the American Enterprise Institute, say it may exceed last year's named Justice Mission 2025 exercise. The Diplomat's analyst sees long-range deployment and cross-theatre command practice, not a blockade rehearsal. I found nothing in today's mainland feed. Initium, publishing outside the mainland, asks why Beijing announced no drill.

THE TELL: no name. My read: the silence is the instrument. In Shenzhen, Beijing can call it routine. In Taipei, with local elections at the end of the month, it reads as rehearsal.

Both sides are auditing the truce. On a 6 November video call, Treasury Secretary Scott Bessent and Trade Representative Jamieson Greer pressed Vice Premier He Lifeng, China's lead trade negotiator. They wanted him to honour the commitments China made at last year's Busan summit on rare earths and farm goods.

Two readouts, one call. Xinhua's account spoke instead of "serious concern" about recent US trade restrictions. China's commerce ministry says it will buy more US farm goods only if reciprocal tariff cuts covering $30 billion of products each way take effect. Against that, Monday's extension of the tariff truce to 10 January looks like a truce with a compliance review attached.

The agenda is widening. In New York, Bessent pitched "de-risk, not decouple" while Chinese ambassador Xie Feng stressed Beijing's red lines. SCMP's sources say Beijing, Washington and Moscow could agree a commitment on controlling AI in nuclear arsenals. Japan's prime minister Sanae Takaichi, whose Taiwan remarks a year ago froze relations, gets no Xi meeting, Kyodo reports via Sinocism.

Go deeper on this section: ClaudeChatGPT

Alibaba's chip IPO and Baidu's peak pricing

THE NUMBER: 560,000. That is how many Zhenwu AI chips T-Head, Alibaba's in-house chip unit, had shipped by April, per Alibaba's announcement. Yesterday I said no company had confirmed a Hong Kong listing. It has now: T-Head filed confidentially on 11 September for an H-share listing, size undecided. The $2–3 billion I relayed stays unconfirmed.

The window behind it. Leiphone's feature credits the listings of domestic GPU makers Moore Threads and MetaX with reopening investors' exits. One quantum-chip startup's valuation rose tenfold in a few months. One headhunter cites an engineer with two years of NPU experience on over 1 million yuan (~$141,000). The piece's thesis: capital runs in weeks, silicon in years.

Baidu priced by the hour. Ernie 5.5 Flash Preview has 280 billion parameters, 14 billion active, a one-million-token context and over 200 tokens a second. List prices run 2 yuan (~$0.28) per million input tokens at peak and 1 yuan off-peak. Output costs 8 yuan (~$1.13) at peak and 4 off-peak, with a limited-time half-price offer. Weights open on 25 November. Peak and off-peak tiers are what a capacity-short cloud does.

Kimi sells hours, not tokens. A TMTPost column on Moonshot AI's hosted agents notes the sandbox costs 0.072 yuan (~1 US cent) an hour, plus token and plugin fees. It says Anthropic's managed agents run $0.08 a session-hour. The sandbox is a rounding error. The argument that matters is the budget line: buyers who compare agents with salaries, not API invoices, stop shopping on token price.

Go deeper on this section: ClaudeChatGPT

Teaching robots pays 180 yuan a day

The floor of the data industry. A 36Kr reporter took a shift as a robot-data collector. Base pay was 180 yuan (~$25) a day, with sensors taped to the fingers and 155 two-minute sets of pinching and lifting props. Software scores each set, and below 80% you cannot leave. Overtime was promised at 30 yuan (~$4) an hour. A month later only base pay arrived, and the dispatcher said no client money had come.

A cushion, not a career. One colleague was a laid-off internet worker. The reporter calls the job a buffer between unemployment and the next job. Xinhua in September quoted an industry source putting part-time collectors near a million. The sensors could be fooled: shaking a bottle instead of flipping it still scored.

The other end of the line. Alibaba's Quark platform has stopped recording human-taught problem-solving videos, per unnamed sources cited by Sina Tech, possibly to make way for Alibaba's Qianwen assistant. Part-time teachers earned nearly 10 yuan (~$1.40) a question on popular subjects. Piece-rate labour trains the machine at one end and loses to it at the other.

Singles Day sold tokens. Tmall made token top-ups a stand-alone category, with all three MiniMax monthly cards at half price. Li Jiaqi, China's top livestream seller, moved 2 trillion tokens in one stream. Reuters found AI adoption surged while festival sales rose only 2.4% from 7 to 31 October. My read: when volume won't grow, the novelty gets the shelf.

A homecoming becomes a loyalty test. SCMP reports that Yuan Junying, a pioneer in the study of how cells die, left Harvard for China. She drew online backlash after defending her US citizenship on winning a major prize in Shanghai. "I've always said my blood is Chinese blood," she said. The row feeds a wider argument over whether senior returnees squeeze out young scientists.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • Huang in Shenzhen: the invitation reversed. Trump said Sunday he would bring Huang to China, per the Times. Sinocism now says Huang is left out. Business Insider, via 36Kr, says he withdrew from two events. He had until 11 November to answer the House request, and I see no reply.
  • T-Head IPO: confirmed, unsized. Alibaba's filing is real. The $2–3 billion figure is not.
  • Anthropic: I was low. Yesterday I wrote $1.2–1.3 trillion. The roadshow range of $70–78 a share implies close to $1.4 trillion fully diluted at the top. Google and Nvidia are each putting in $10 billion through private placements.
  • The purge: a scientist's page vanishes. On 10 November I noted General Zhang Shengmin, a Central Military Commission vice-chairman and former PLA anti-corruption chief, calling for joint military-civilian graft-fighting. SCMP now reports that PLA cybersecurity researcher Guo Shize, 57, elected to the Chinese Academy of Sciences in 2023, has been removed from its website. Neither institution has commented. A deleted page is not a charge.
  • The car slump: Tesla and Mercedes join FAW-VW. Dongchedi rankings show Model Y retail at 7,807 in October, against 50,359 in September. Mercedes chief Ola Källenius says China's market fell more than 20% and luxury 30%. Mercedes held prices and gave up volume.

Go deeper on this section: ClaudeChatGPT

The river · 34

China-focused labs buy US-built task environments and grading rubrics for model post-training; the strategic question is whether this is ordinary data commerce or a consequential channel for capability transfer.

More
  • The article says US vendors sell Chinese labs data and reinforcement-learning environments for domains including coding, finance, law, cybersecurity and biology.
  • An environment can include a simulated software tool and a rubric that rewards or penalizes model actions, letting developers train tool use through repeated attempts.
  • The distinction from distillation matters: vendors provide training inputs and task setups, not necessarily a copy of another model’s outputs.
  • The author argues that post-training can account for major capability gains between releases that share the same pretrained base model.
  • The policy case depends on what is being transferred, who receives it and whether controls can distinguish routine commercial data from unusually sensitive capability-building tools.

Hong Kong sentenced Joshua Wong to another 10 years and Macau sentenced Au Kam San to 12 and a half years, extending the use of national-security law against political dissent.

More
  • The report says Wong received the additional Hong Kong sentence for conspiracy to collude with foreign forces; he had already been serving a prison term.
  • Au, a former Macau legislator and pro-democracy activist, was sentenced in a closed-door trial for subversion and other charges.
  • The cases show national-security prosecutions reaching across both special administrative regions, while leaving little space for organized democratic opposition.
  • The report notes criticism from governments and human-rights groups, which have called for the sentences to be overturned.
  • For Hong Kong, the added sentence means Wong remains imprisoned well beyond his previously expected release date, further sidelining a prominent organizer.
  • The broader signal is the long-term political cost of activism: prosecution can remove organizers from public life for years, not merely constrain a particular campaign.
  • The article reports that Wong pleaded guilty to the charge; it also describes disputes over evidence and whether pre-law conduct figured in the case.

Baidu’s preview model pairs sparse attention and routed experts with a million-token context window, targeting fast coding and multi-step work while promising open weights later this month.

More
  • The model has 280 billion total parameters, with 14 billion activated per token, and supports a context window of up to one million tokens.
  • Baidu says a new mixed sparse-attention design, multi-router mixture of experts, and multi-token prediction reduce compute and memory costs while enabling decoding above 200 tokens per second.
  • The model is trained for tool-heavy workflows including software development, Blender and CAD use, document work, and video creation; the article describes dedicated simulated workspaces and reference files for professional tasks.
  • The preview is available through Baidu’s cloud model platform, with time-of-day API pricing and a promotional discount; Baidu says the model weights will be released to the global open-source community on November 25.
  • These are company-reported performance and training claims; the article does not provide independent benchmark results sufficient to establish broad superiority.

China, Russia and the US have discussed adding controls on AI in nuclear arsenals to leader-level talks around APEC, putting strategic-risk governance alongside broader diplomacy.

More
  • The report says the three countries have spent weeks negotiating the agenda and protocol for leaders’ meetings on the sidelines of the Shenzhen APEC summit.
  • The proposed commitment concerns AI in nuclear arsenals; the report does not specify whether it would be binding or what systems it would cover.
  • This is a narrow but consequential governance topic: safeguards against AI affecting nuclear command decisions could be politically easier to discuss than broader limits on military AI.
  • The talks are part of a wider agenda, with Ukraine and Korean denuclearization also reportedly under discussion.
  • Whether leaders agree to language, and whether it establishes a durable process rather than a summit statement, will determine its significance.

A reported visit to a Chinese robot-data site finds low-paid temporary workers repeating hand motions for training data, revealing a labor layer behind embodied-AI development.

More
  • The reporter describes workers performing repetitive object-grasping tasks while sensors record their motions; a stated daily target is 155 sets of actions.
  • The work is organized through subcontractors between robot or model companies and the temporary labor force, with little visibility into where the data ultimately goes.
  • The article reports a daily wage of 180 yuan (roughly $25), and describes workers who are unemployed or using the job as a stopgap.
  • Data collection is not an incidental support function: physical demonstrations are a scarce input for training systems to manipulate varied objects.
  • The account suggests a new low-wage task category is emerging around AI, with performance targets and monitoring resembling factory work.
  • The article is an individual field report, not a representative labor survey; it leaves open how widespread these conditions are and what worker protections apply.

Alibaba is preparing a Hong Kong listing for T-Head, a broad in-house chip designer whose commercialization gives investors a rare way to value Chinese hyperscaler silicon separately.

More
  • Alibaba says T-Head is expected to remain a subsidiary after the proposed H-share offering; the size of the deal and Alibaba's eventual stake are not yet set.
  • T-Head spans AI accelerators, server CPUs, interconnect, network cards and storage controllers, a wider portfolio than a single accelerator bet.
  • The article reports more than 560,000 Zhenwu AI chips shipped by April and more than 650 customers by August, while Alibaba has not disclosed standalone T-Head revenue.
  • Separating the unit could give it an independent financing channel and let investors distinguish chip economics from Alibaba's cloud business, though control is expected to stay with the parent.
  • The filing arrives as several Chinese internet groups are building internal silicon for compute supply and workload-specific optimization.
  • A public listing will test whether reported shipments translate into recurring external sales, credible margins and a roadmap competitive beyond captive demand.

China supplied 21% of Brazil’s imported medical devices in the first half of 2026, edging past the US as Chinese vendors move beyond consumables into imaging and monitoring equipment.

More
  • Brazil imported $696.5 million in medical devices from China in the first half of 2026, compared with a 20% share for the United States and 12% for Germany.
  • China’s share rose from 12.4% in 2025, when the United States was still the leading supplier; the shift comes amid growing Brazilian healthcare activity and equipment demand.
  • Chinese products are particularly prominent in medical imaging, hospital monitoring and support equipment, while US suppliers retain stronger positions in in-vitro diagnostics and several high-end clinical categories.
  • The article describes United Imaging’s planned 535 million-real factory investment in Brazil, intended to manufacture MRI and cancer radiotherapy equipment and create 200 to 500 jobs.
  • The strategic test is whether Chinese suppliers can convert equipment placements into durable service, consumables and clinical relationships, rather than relying on price and shipment volume alone.

China's official account of the call stresses continuity, cooperation and summit coordination; a separate report describes the same pre-summit contact.

More
  • Wang Yi said Beijing's policy toward the United States remains stable and called for communication on disputes and sensitive issues.
  • The account links bilateral diplomacy to preparations for APEC and the G20, and describes the exchange as positive and constructive.
  • The report provides Beijing's public framing, not a detailed account of Rubio's response or any specific agreement.

A Rhodium Group estimate puts 2026 AI infrastructure spending by 13 Chinese firms at 932 billion yuan (roughly $130 billion), while Alibaba, Tencent and Baidu spent more than their combined free cash flow in the first half.

More
  • The report distinguishes capital expenditure from AI-only spending and notes that negative free cash flow is not the same as operating losses; the signal is how quickly infrastructure investment is consuming cash.
  • Alibaba illustrates the timing problem: cloud and AI revenues are growing, but equipment purchases, power, maintenance and depreciation arrive before utilization and customer revenue fully catch up.
  • Tencent and Baidu face different pressures: Tencent is committing large sums to secure compute, while Baidu must fund AI growth as its established advertising business weakens.
  • For model companies, rising API prices do not guarantee better margins; the cited Chinese labs saw gross margins fall in the first half of 2026.
  • The key test is not announced capacity or model growth, but whether customers use new infrastructure enough to cover its financing and operating costs.

A Chinese company says its L4 tractor system completed deep-plowing work in Xinjiang without an onboard driver, a test of autonomy in the long, demanding work windows of large farms.

More
  • The company reports plowing deeper than 30 centimeters at 11.8 kilometers per hour; it also cites a 1.2-centimeter average lateral error and more than 3.2 kilometers without a human takeover.
  • The system adds an autonomous-operation kit to a production tractor rather than redesigning the whole vehicle, combining perception, control, task management and remote monitoring.
  • The report highlights sensor time synchronization as an important engineering hurdle: lidar, cameras and inertial sensors must provide aligned data for fusion and mapping.
  • Farm equipment is part of the control loop: the tractor’s autonomy is limited if implements cannot communicate status or respond reliably to commands.
  • The company also describes a telemetry platform and plans to open a direct-sales outlet, but commercial scale and all-season reliability remain unproven.

Google has expanded access to Gemini 4 Carbon; Chinese coverage emphasizes agent benchmarks and demos while noting its specialization for difficult tasks.

Also: InfoQ China

More
  • Google says Carbon leads selected agent workflow benchmarks and costs less per task than the next-best external model in those tests.
  • The article also reports that Claude Fable 5.2 scores higher on some broad and software-engineering evaluations, showing that rankings depend on workload.
  • Carbon is intended for difficult debugging, scientific research and extended agent work; Google says its Argon model remains better for most routine tasks.
  • A multi-agent feature in Antigravity lets agents split work across isolated environments and coordinate review.
  • Google's demos demonstrate capability, but benchmark selection and company-provided examples do not establish reliable performance on typical production codebases.

Chinese robotics company Daxiao describes a hotel-laundry system that rehearses tasks in a physics-aware digital twin, then executes learned skills locally instead of waiting on a cloud model at every step.

More
  • The company says its Mundo world model represents physical properties such as mass and friction, allowing simulated practice in a hotel-like environment.
  • A general cloud model is assigned planning and strategy work, while a local robot model executes stored skills at lower latency.
  • The company reports that a complete task took more than 30 minutes with cloud-only control and 18.6 seconds after skill-based execution; the excerpt provides no independent test protocol.
  • A task scheduler reportedly decomposes the workflow into atomic actions and higher-level steps, adapting sequences to room location and equipment availability.
  • The architecture addresses a real robotics constraint: high-level reasoning may be useful for exceptions, but low-latency manipulation cannot wait on remote inference continuously.
  • Whether the system generalizes beyond a controlled hotel environment depends on recovery from occlusion, shifted objects and other unmodeled physical events.

A personal essay connects Mongolia’s school-language changes with a new national-unity law that elevates the state language in classrooms and public settings.

More
  • The essay says the law, effective July 2026, makes the national language the basic language of instruction and gives it precedence in public settings where minority languages are also used.
  • It contrasts formal guarantees of minority-language study with the loss of some subjects previously taught through Mongolian.
  • The author recalls Inner Mongolia’s 2020 curriculum changes, which shifted several subjects to centrally compiled textbooks and Chinese-language instruction while leaving Mongolian-language classes in place.
  • The policy debate is not simply whether students should learn the national language; the essay focuses on whether minority languages retain the ability to carry a full education.
  • The author acknowledges that Chinese opened access to wider educational and cultural opportunities in their own life, while questioning what is lost when Mongolian’s role narrows.

AIxCoder has released a small model that summarizes and prunes coding-agent histories, aiming to reduce repeated context input while preserving decisions, tool results and task state.

More
  • The model classifies history into content to retain, tool output to shorten and redundant records to remove, then creates a structured task summary.
  • The company says it can run locally on a single 40GB A100 and compresses historical context to an average of about 26% of its original length in its target tasks.
  • AIxCoder reports task completion comparable to DeepSeek-V4-Pro on Terminal-Bench after integrating the compressor with OpenCode; this is a vendor-reported result.
  • The architectural tradeoff is consequential: aggressive compression reduces cost but can discard constraints or debugging evidence that an agent needs later.
  • The model’s open release allows teams to inspect and adapt the compression layer rather than sending long coding traces to a third-party service.
  • Independent tests across different repositories and agent frameworks will determine whether the reported savings preserve reliability.

A Chinese analysis models how local permitting resistance and new state policies could constrain US data-center expansion and future frontier-model training.

Also: TMTPost

More
  • The article estimates 230 to 340 gigawatts of planned but unbuilt US data-center capacity, while warning that pipeline figures include speculative queue requests.
  • In its extreme scenario, capacity reaches roughly 120 to 150 gigawatts by 2028 and then stops growing, rather than reaching the far larger levels implied by all announced plans.
  • The near-term effect would be limited because major facilities already under construction could support models through 2027; the pressure would fall on projects needed for later, larger training runs.
  • The report links opposition to electricity costs, land, water and local politics, showing that AI infrastructure is becoming a community and regulatory issue, not just a cloud procurement decision.
  • Its scenario should not be mistaken for a forecast: the source acknowledges substantial uncertainty in planned capacity and excludes possible capacity relocation.

A Chinese influencer demonstrated an AI coding tool during Double 11, presenting it as a way for non-programmers to turn rough ideas into working documents and software.

More
  • The demonstrations included producing a market report, adapting a long document into slides and building a voice-input utility from a sketch.
  • The article emphasizes that the tool can inspect multiple files, configure environments and continue a task after interruption; these are vendor demonstration claims.
  • The wider signal is a shift from selling raw compute quotas to selling an application that promises a finished result.
  • The performance of a live demo does not establish that the tool can safely maintain complex software or replace engineering review.

US officials urged China to meet agricultural and rare-earth commitments, while Beijing told Washington to implement reciprocal tariff cuts before buying more farm goods.

More
  • The exchange shows that both governments are treating existing deals as leverage in the run-up to a leaders' meeting, rather than assuming commitments are self-enforcing.
  • Agriculture and rare earths are different pressure points: one matters to US producers and politics, the other to Chinese supply-chain leverage.
  • The supplied report describes competing demands, not a settled bargain or a confirmed change in trade flows.

A hands-on article shows ERNIE 5.5 Flash Preview generating a browser game and animations, while noting that extended tool-driven demos require repeated design and testing.

More
  • The model is reported to have created a 116-second animation using Python and a playable browser-based game.
  • The story argues that coding capability increasingly means completing tool-mediated projects rather than generating snippets.

A Chinese chip entrepreneur reportedly declined 50 million yuan (about $7 million) to start a company, while another founder describes pressure from rapid fundraising cycles.

More
  • The article says some hot chip startups saw valuations rise sharply within months, even though chip design, fabrication and customer validation take years.
  • A founder working on an inference processor says limited time for architecture research was one cost of fundraising pressure.
  • The boom is concentrated in AI compute and specialized inference hardware, while other semiconductor categories attract less investor attention.
  • The story highlights a funding-market mismatch: investors price optionality quickly, but engineering risk resolves slowly.

Ahead of the Shenzhen summit, both governments appear interested in keeping a fragile detente intact; the analysis frames domestic pressures as a constraint on how far either side can move.

More
  • The report places the APEC gathering on November 18–19 in Shenzhen, alongside a planned Xi-Trump meeting.
  • It says the leaders have previously used the phrase “constructive strategic stability” to describe the relationship.
  • The article argues both governments have domestic reasons to avoid a rapid deterioration in ties.
  • The “low-hanging fruit” framing suggests limited, practical agreements are more plausible than resolving core disputes.
  • This is analysis of incentives and timing, not a report of an agreed summit outcome.

A Japanese media account describes a failed effort to answer Beijing's pressure after Prime Minister Sanae Takaichi's remarks on a Taiwan contingency.

More
  • The report traces the episode to a diplomatic meeting in Beijing a year after Takaichi's comments froze relations.
  • The headline frames the outcome as a successful Chinese coercive tactic, but the excerpt does not provide enough detail to independently assess the negotiation.
  • The episode illustrates how Taiwan-related language can become a lever in China-Japan diplomacy.

US Secretary of State Marco Rubio raised a possible trilateral summit with China and Russia in a call with Wang Yi, adding a new diplomatic question ahead of the Shenzhen APEC gathering.

More
  • The US State Department readout described agreement on the value of leader-level conversations about peace and denuclearization.
  • The proposal would bring Washington, Beijing and Moscow into a format that could complicate existing alliances as well as create room for diplomacy.
  • The supplied report says a trilateral meeting has been discussed in media coverage but does not confirm that one is planned.
  • For Beijing, the idea would offer another venue to present itself as a major diplomatic actor; for Washington, it would test whether coordination can extend beyond bilateral ties.
  • The timing, on the eve of the APEC summit, makes it part of a broader effort to manage several strategic relationships at once.
  • The next signal is whether either government confirms a concrete format or treats the call as exploratory.

A Chinese-American biologist who left Harvard for a role in China faced online backlash after defending her US citizenship, amid concern that celebrated returnees may crowd out younger local researchers.

More
  • The scientist, Yuan Junying, is described as a pioneer in cell-death research and drew attention after receiving a major science prize at a Shanghai forum.
  • Her public remarks about Chinese identity and US citizenship became a focus of online criticism, according to the report.
  • The argument reaches beyond one appointment: it reflects tension between recruiting internationally trained senior scientists and building stable advancement paths for younger researchers in China.

MiniMax has open-sourced H3-Image, which the report says can generate a 4K image in about 10 seconds and accept up to 16 reference images; the article’s own tests emphasize instruction following.

Also: New Wisdom

More
  • The model is described as supporting 4,096-by-4,096 output and complex compositions built from multiple reference images.
  • The report says the model handled dense poster text and a multi-view concept design in the outlet’s hands-on tests.
  • A 16-image test combined people, clothing, color palettes, typography and layout references into one poster task.
  • The article makes strong quality claims based on a small set of editorial examples, not standardized evaluation.
  • Open weights make it easier for developers to test deployment and fine-tuning, but licensing and compute requirements are not detailed here.

Chinese-focused coverage warns that US hawks could turn espionage allegations involving Huang into a bilateral issue, while much US reporting has centered on the board and personal-conduct story.

Also: Techmeme

More
  • This is a collection of related articles, not a new report about the allegations.
  • The framing highlights the risk that unverified claims about Chinese intelligence could be used in US political debates over Nvidia and China.
  • The excerpt also points to Chinese experts’ warnings that the controversy could worsen bilateral tensions.
  • No evidence in the supplied text establishes that Nvidia secrets were compromised.

Wang Yi told Cuba’s foreign minister that Beijing would oppose sanctions and foreign interference, days after Washington penalized two China-linked firms over alleged Cuban military supply.

More
  • The ministry’s readout presents China’s position as support for Cuban sovereignty and opposition to sanctions, without detailing material assistance.
  • The timing links the diplomatic pledge to US action against Chinese state-linked firms, turning Cuba policy into a point of bilateral friction.
  • The source reports the Chinese readout but does not provide the US response or the firms’ account.

A Chinese hands-on review found Qwen4-Max preview capable at interactive coding and media generation, but reported mismatched subtitles, physical inconsistencies and long generation times.

More
  • The preview release did not include published parameter counts or benchmark results, leaving independent task tests as the main early evidence.
  • The reviewer reports a browser-based train-crash simulation with adjustable physics and sound, plus a generated animated GIF based on an uploaded image.
  • In a longer dialogue-driven GIF, character motion was convincing but dialogue attribution and subtitle timing were wrong.
  • The article also reports a roughly 14-minute generation for one multimodal task, a reminder that output quality and completion latency are separate dimensions.
  • These are anecdotal tests on a preview model, not a controlled comparison against competing systems.

China lost more than 60,000 kindergartens from 2021 to 2025, and parents describe a rapid shift from competing for places to watching schools shut down.

More
  • The account connects falling enrollment to lived changes in childcare, school admissions and the fortunes of private kindergartens, rather than treating closures as a statistic alone.
  • One family’s experience contrasts the scramble to secure maternity and preschool slots for a 2016 child with the ease of arranging care for a younger sibling.
  • The reversal has consequences beyond families: operators and teachers face closures or shrinking enrollment, while education capacity and jobs must adjust to fewer children.

US Treasury Secretary Scott Bessent advocates de-risking rather than decoupling, while China's ambassador emphasizes sovereignty and red lines ahead of a leaders' summit.

More
  • The exchange places economic risk reduction alongside political limits on what either side will accept.
  • The report says both speakers linked bilateral ties to the planned meeting between Trump and Xi, but the excerpt identifies no concrete trade agreement.
  • The paired language captures the narrow space for economic engagement amid unresolved strategic disputes.

Snowflake AI Research describes context management and upfront business knowledge as ways to make enterprise agents cheaper and more reliable.

More
  • The proposed Harness loads tool definitions only when needed, combines independent calls and compresses large tool outputs instead of carrying them through every model turn.
  • The second path is to supply business rules and context before execution, reducing exploratory searches and limiting ambiguity about what the agent should treat as production data.
  • In one internal SQL-repair test, CoCo using Sonnet 5 reportedly passed 86% of trials versus 72% for Claude Code using Opus 5, with reported cost 33% lower.
  • Those figures apply to Snowflake's specific test and configuration, not a general guarantee that a Harness improves every agent workload.

Chinese coverage describes Alibaba’s proposed T-Head spin-off alongside the Anthropic IPO story, highlighting a domestic chip company seeking a separate public-market valuation.

Also: 36Kr

More
  • Alibaba says T-Head has filed confidentially for a Hong Kong main-board listing and is expected to remain a subsidiary after the offering.
  • The listing size and any change to Alibaba’s ownership are not yet determined, and the transaction still requires regulatory and exchange approvals.
  • The roundup’s wider focus is global AI financing, but T-Head is its main China-specific corporate development.
  • A separate public listing could improve visibility into Alibaba’s chip business without necessarily changing operational control.

Chinese coverage reads record chip-led exports in South Korea and Taiwan as evidence that Asian trade competitiveness is shifting toward AI supply-chain bottlenecks, while stressing how concentrated and price-sensitive the boom is.

More
  • The article reports South Korean October exports of $115.3 billion and Taiwan exports of $96.8 billion, driven heavily by semiconductors rather than a broad manufacturing rebound.
  • It cautions that low comparison bases and memory-price increases inflate headline growth; it estimates underlying chip-export growth more plausibly at roughly 30-40%.
  • China's chip export value rose 137% year over year to about $39.6 billion even as export volume fell 2.5%, suggesting price and product mix drove much of the increase.
  • The supply chain concentrates value in Taiwan's advanced foundries, Korean HBM suppliers and regional server assembly, making Asian exports unusually dependent on a small set of US cloud buyers.
  • This is a strong demand signal but also a concentration risk: capacity decisions based on peak growth rates could overshoot if hyperscaler spending cools.

China’s first AI-assisted feature film took in about 1.68 million yuan (roughly $240,000) in 11 days, turning a showcase for Shengshu’s video model into a stress test of its commercial claims.

More
  • The film’s weekday attendance reportedly fell below one viewer per screening, a conspicuous result for a project positioned as proof that generative video can serve professional production.
  • Shengshu’s Vidu is strongest, according to the article, at maintaining consistency from reference images; longer narrative sequences and live-action faces remain harder cases.
  • The company has promoted fast, low-cost video generation and reported early recurring revenue, but a weak feature-film result makes the gap between short controlled clips and sustained storytelling more visible.
  • The timing is notable because Shengshu is preparing for a Hong Kong listing in a less exuberant market for AI companies, where investors increasingly want evidence of durable revenue.
  • The film is not a controlled technical evaluation of Vidu; production choices, story quality, distribution and audience response could all contribute to the result.

A Chinese account of Singles' Day traces a shift from growth-era discount fervor to cautious, self-directed spending, as platforms stop publishing headline sales totals.

More
  • The article contrasts the early years, when the shopping festival cultivated mass participation and brands reported dramatic sales, with today's more restrained consumers, who are less inclined to stockpile goods just to qualify for discounts.
  • It describes a move toward smaller purchases tied to personal enjoyment and emotional comfort, rather than conspicuous deal-hunting. The piece presents this as a change in consumer mood, not as a quantified survey finding.
  • Platforms began pulling back from publishing gross merchandise volume figures in 2022, according to the article, and Tmall dropped its media open day by 2025. The vanishing scoreboard marks a retreat from the spectacle that once helped make Singles' Day a national event.
  • The article also points to merchants' difficulty judging whether promotions are profitable amid traffic costs, discounts and returns, suggesting the festival's growth-era playbook has become harder to sustain.