Bloomberg reports that DeepSeek raised roughly 1 trillion yuan (about $140 billion) from investors including CATL and Tencent, a record for China's AI sector.
Also:
36Kr
More
- The report says DeepSeek increased its fundraising target twice, from 500 billion yuan to 800 billion yuan and then to about 1 trillion yuan (roughly $140 billion).
- The financing would give the lab resources for data-center infrastructure and continued development of open models for Chinese and overseas users.
- A large private round could help fund a future listing, but the report says details remain undisclosed and does not establish that an IPO is imminent.
- DeepSeek's open-model strategy has contrasted with some US labs' more closed commercial approach; the new capital could test whether openness can coexist with infrastructure-scale spending.
- The amount is extraordinary and should be treated as a reported figure until confirmed by the company or filings.
The company says its Spark X2.6 combines a million-token cloud model with a 4B edge model, using a domestic compute stack and routing work between device and cloud.
More
- The cloud model supports a one-million-token context window; published prices are 0.9 yuan (about $0.13) per million input tokens, 0.08 yuan (about $0.01) for cached input, and 3.6 yuan (about $0.50) per million output tokens.
- iFlytek says its domestic training cluster cut one long-context training run from 4.5 months to 1.2 months, and that its four-layer KV-cache design raised cache hit rate from 63 percent to 95.7 percent.
- The edge-cloud system routes tasks among on-device, cloud, and split execution; the company says training improved routing accuracy from 81.2 percent to 96.1 percent.
- The 4B edge model is positioned for PCs, cars, and robots, while local processing and redaction before cloud escalation are presented as privacy and cost controls.
- These are vendor claims and benchmark results; the strategic point is a full-stack domestic compute narrative tied to lower-cost deployment, not proof of parity across all workloads.
At a major commemoration, Xi linked the Red Army's victory to correct Party leadership and a strong central core, shortly before a key Party plenum.
Also:
SCMP
More
- The newsletter says Xi's speech made Party leadership the first and longest of four lessons from the Long March.
- It contrasts the speech with the 2016 anniversary address, which put greater emphasis on ideals and faith.
- The political message is that unity around central leadership is being presented as a prerequisite for overcoming present-day challenges.
- The event drew senior Party leaders and about 3,000 attendees, according to the newsletter.
AstraZeneca says it will sell in the US a China-developed cancer drug after some Americans traveled to Shanghai to obtain it sooner.
More
- The story exposes a gap between regulatory approval and commercial availability: the drug was approved by the US FDA but was not readily available there.
- Chinese coverage foregrounds patients traveling to China and pressure on AstraZeneca, making access and distribution—not just drug discovery—the story.
China activated a national level-four earthquake response after a reported magnitude-7.0 quake in sparsely populated, high-altitude Nagqu; central officials dispatched a team to guide rescue and monitoring.
More
- The emergency ministry says the quake struck Baqing County at a depth of 10 kilometers, with the epicenter near 4,800 meters above sea level.
- The national earthquake-relief headquarters and the emergency ministry initiated the level-four response and sent a team to the affected area.
- The account emphasizes the government’s rapid dispatch and coordination; it gives no casualty or damage figures.
- The sparse population may limit immediate human impact, but remoteness and altitude complicate rescue and assessment.
- This is an initial official response, not a final damage assessment; the scale of the event depends on verified reports from the area.
A Zhejiang University team reported high-purity silicon-28 gas and a four-inch epitaxial wafer, extending domestic isotope work into a key substrate step for silicon quantum devices.
More
- The team reported silicon-28 abundance of 99.99576% in its gas and 99.9937% in the wafer’s epitaxial layer; chemical purity of the gas was reported separately at 99.99998%.
- Silicon-28’s nuclear spin is zero, making it a promising substrate for silicon quantum devices that are sensitive to nuclear-spin noise.
- Earlier domestic work described in the article achieved isotope enrichment; the new report adds conversion into a semiconductor-grade epitaxial layer.
- The results establish material measurements, not a working quantum processor: coherence time, gate fidelity, interface defects and device fabrication remain untested in this report.
- The useful milestone is a domestic path through both gas production and wafer growth; whether it reduces dependence on external material supply will depend on reproducible scale and device performance.
A ChinaTalk essay examines the emerging labor and infrastructure behind egocentric data collection, including the challenge of capturing dexterous hand movements.
More
- The essay places embodied intelligence among China's national priorities and says more than 150 humanoid-robotics firms and 40 state-owned data centers were reported by the end of 2025.
- It cites a report that Chinese embodied-AI startups had attracted 50 billion yuan in investment deals by September 2025 (roughly $7 billion).
- Unlike clicks and likes, human-hand motion must be deliberately recorded and demonstrated, making data collection a physical labor and coordination problem.
- The story points to a deeper industrial question: whether China's large robotics ecosystem can turn demonstrations and data facilities into reusable training datasets, rather than isolated showcases.
Recursive, co-founded by Chinese researcher Tian Yuandong, released Flint, an open-weight MoE coding model that the company says can generate more than 125 tokens per second on an M4 Pro MacBook Pro.
More
- Flint has 39.8 billion total parameters but activates about 2.5 billion per token, a sparse design intended to reduce local inference cost.
- Recursive reports scores of 54.9% on Terminal-Bench 2.1, 89.7% on SWE-Bench Pro v2 and 74.2% on SWE-Bench Verified; those are company-published benchmark results.
- The model was built from scratch through tokenizer development, pretraining, supervised fine-tuning and reinforcement learning, according to the report, and is released under Apache 2.0.
- The local-first target is codebase analysis, bug fixing, feature work and tests without sending the full workload to a cloud service; the tradeoff is that complex jobs may need more explicit task decomposition.
- The story also connects the launch to Tian’s stated interest in recursive self-improvement, but the released model itself is a coding model, not evidence that the broader research goal has been achieved.
At Unitree's tenth-anniversary event, Wang Xingxing moved his humanoid-robot forecast forward and argued that hardware scale may matter more than owning the strongest model.
Also:
TMTPost, Jiemian, IT Home
More
- Wang put a possible physical-AI breakthrough within one to three years, down from a broader timeline he gave in August.
- The article says Unitree's G1 has produced roughly 10,000 to 20,000 units, while Wang argues mass adoption would require far larger manufacturing capacity.
- He expects frontier-model providers to supply increasingly capable robot brains, leaving robot makers to compete on reliable, affordable bodies and real-world integration.
- The company plans to develop its own models and a loop that searches research, writes control code, trains in simulation, tests on robots, and feeds results back.
Beta Infinite's BetaMem system aims to track how objects and environments change over time, with the company reporting top results on four memory benchmarks.
More
- BetaMem separates a current world-state graph, an append-only event log, and a limited archive of source images and audio, linking conclusions back to evidence.
- That architecture addresses a central failure mode in embodied systems: new observations can overwrite old states, while stale or incorrect memories can contaminate later decisions.
- The company reports that a 9B open model achieved leading results across four benchmarks, but the supplied report does not provide independent replication.
- The system's value depends on memory management as much as retrieval: deciding what to retain, when to verify it, and how to correct an error are central design problems.
- The research contrasts with narrower mapping approaches by treating memory as an auditable, revisable judgment rather than a transcript of everything a robot has seen.
AI comic-drama production costs have fallen sharply while output has surged, yet interviewees say monthly revenue has barely grown and platform payouts are shrinking.
More
- One producer says an 80-minute show fell from about 100,000 yuan (roughly $14,000) to 40,000 yuan (roughly $5,600) in production cost over a year.
- The same account says one person can now produce about three shows a month, versus a ten-person team making one, but the producer's monthly revenue stayed around 20-30 million yuan (roughly $2.8-4.2 million).
- As labor costs fall, token costs have become a larger share of production expenses, shifting the scarce input from editing capacity toward compute, story selection, and creative judgment.
- Platforms are encouraging individual creators and studios, weakening the old advantage of large production houses; the labor transition is real even as interviewees' forecasts remain speculative.
Beijing reportedly offered export-control concessions in exchange for the EU easing sanctions on a Chinese textile maker accused of supplying sensitive technology to Russia.
More
- The proposed exchange links EU sanctions on Yusha Group to China's promise that it will not ship banned goods to Russian military-linked buyers.
- Beijing also reportedly offered to remove German firm Sindlhauser Materials from its own export-control list.
- The case turns sanctions relief into a negotiated trade: each side is using market access and controls as leverage.
- The report relies on people familiar with the talks; the supplied account does not establish that the deal is final.
The Qwen team describes FlashQLA, a GPU kernel library for gated delta networks that addresses sequential dependencies and memory traffic in linear attention.
More
- The article positions linear attention as a way to reduce sequence-length scaling from quadratic toward linear, while noting that recurrent state updates create a parallelism bottleneck.
- FlashQLA targets Gated Delta Network layers in hybrid models, rather than replacing softmax attention everywhere; the article describes Qwen's three-to-one layer mix.
- A WY-representation transformation groups token updates into chunks, allowing more work to use GPU matrix units instead of a token-by-token recurrence.
- The implementation combines sequence chunking, warp specialization, hardware-friendly algebra, and reuse of sliding-window properties to reduce redundant memory movement.
- The engineering tradeoff is explicit: linear attention is efficient at long context but compresses history into a limited state, so occasional softmax layers help preserve long-range recall.
China Telecom’s Xing4.0-29B-Intent combines intent classification with structured extraction, producing schema-bound JSON for systems such as customer-service routing and smart-home controls.
More
- The model and 2.75 million training examples are being released with weights and an inference image, according to the company.
- It uses a unified schema-driven parsing framework to return both an intent category and the relevant entities or parameters.
- China Telecom reports a 58-point aggregate score across four extraction tasks, above the best general model it tested at 49; the evaluation is company-run.
- The model is already used in the company’s customer service and smart-home products, but independent evaluation and performance on changing schemas remain open questions.
A Chinese report says Silicon Valley vendor Maxinsights supplies large egocentric-video datasets to robotics labs and entered China this year, spotlighting human video as a scalable training input.
More
- The article attributes more than two million delivered hours of data, over 1.5 million hours with hand tracking and language labels, and collection across more than 2,000 environments to Maxinsights.
- It says Dyna-2 used one million hours of first-person human video and that prediction error fell about 12% as training data scaled from 1,000 to one million hours; the vendor's role and figures are reported claims, not independently established here.
- First-person video offers varied examples of human hand-object interaction without requiring every example to be collected on a robot, but the report does not show how reliably that data transfers to real-world robot control.
- The article says Maxinsights entered the Chinese market in March but gives no detail on local customers or revenue.
Less than two weeks after a deal eased a trade dispute, European officials tout collective pressure while Beijing rejects the idea it was forced to yield.
More
- The disagreement is over the political meaning of the recent agreement, not merely its terms: Brussels and Berlin see leverage, while Beijing resists that framing.
- The article says EU officials are already considering further trade restrictions, so the short-lived rapprochement may not prevent another round of pressure.
- The dispute shows how both sides compete to define concessions for domestic audiences as they negotiate trade and industrial policy.
Chinese coverage says AI helped identify two severe PHP memory-management vulnerabilities, prompting several governments and organizations to take websites offline before patching.
More
- The report says the flaws could enable unauthenticated remote code execution and affected a broad range of PHP installations.
- It traces the discovery to AI-assisted code review and says a reusable search pattern helped researchers find a second issue.
- The article frames AI as reducing the cost of finding bugs that previously required extensive manual work; independent technical validation remains important.
QbitAI reports Ant’s claim that LingDT-Decision-1.0 beats several decision models on selected tests and reduces costs in enterprise workloads.
More
- The model comes in 4B and 27B sizes and targets structured classification, tool selection and risk decisions rather than open-ended text generation.
- Ant says native multi-select and ordered decisions address compound intents and task dependencies that a single-label classifier can miss.
- The benchmark results and business gains are vendor claims; this account is a second report on the same release, not independent validation.
CATL reported strong nine-month revenue growth but weaker third-quarter profit growth and falling margins, as price pressure, product mix and automakers' search for battery alternatives test its bargaining power.
More
- The article reports nine-month revenue of 442.1 billion yuan (about $62 billion) and net profit of 64.5 billion yuan (about $9 billion); third-quarter profit growth slowed to 14.5%.
- Gross margin fell for a third consecutive quarter, while operating costs grew faster than revenue, according to the article's account of the results.
- Storage batteries are an increasing share of shipments, but the article describes that market as price-sensitive and says the shift has not translated into proportional profit growth.
- Automakers are developing alternative supply options, including in-house battery capacity and greater control over battery-management systems, which could strengthen their negotiating position even if CATL retains high market share.
- CATL is pursuing adjacent markets such as data-center power and battery swapping; these could diversify demand but require different delivery models and may not reproduce automotive margins.
A Shanghai research team says a new interconnect design takes single-photon detection beyond the previous 32-channel ceiling, a component-level advance that could support higher-bandwidth photonic quantum networks.
More
- Single-photon detectors are key components in photonic quantum technologies, where they register individual photons carrying quantum information.
- The reported design supports more than 100 detection channels while maintaining accurate detection, compared with a cited prior maximum of 32.
- Increasing channel count could improve the scale and throughput of photonic quantum systems, but a detector alone does not establish a practical quantum internet.
- The excerpt does not identify the detector’s performance under field conditions, its fabrication yield or whether it has been integrated into a network demonstration.
Ant Group's LingDT model makes structured choices directly rather than generating text first, with reported gains in latency and token use.
Also:
Leiphone
More
- The model comes in 4B and 27B versions and targets classification, intent detection, tool selection, and risk decisions rather than open-ended text generation.
- Its key architectural distinction is direct selection among candidate outcomes, avoiding a generate-then-parse step in fixed-choice workflows.
- Ant says its 27B version led a weighted evaluation across 23 open benchmarks; the results are company-reported, not an independent comparison.
- The system adds native multi-select and dependency ordering, representing multi-step tasks as a directed acyclic graph.
- Ant's own tests report lower token consumption and faster handling in customer service and financial workflows, but real deployments still need business rules and permission checks.
Alibaba's glasses combine low-power sensing and multiple chips with access to its travel, shopping, and payment services, making the device an entry point to its agent ecosystem.
More
- The N1 Pro uses separate chips for imaging and compute, routine tasks, and low-power visual understanding, a design intended to balance persistent sensing with battery life.
- The glasses add eye tracking and health monitoring, while Alibaba's pitch is that context from the wearer's surroundings can make its agent more useful.
- The strategic moat is less the eyewear itself than access to Alibaba services and consumer data from mapping, travel, payments, and shopping.
- The product starts at 2,218 yuan (roughly $310); the companion ear device starts at 1,099 yuan (roughly $155), positioning the launch as an ecosystem bet rather than a standalone hardware play.
A dispute over the Huiyuan name has produced competing juice brands: one company says it will stop using the familiar mark, while another continues selling it, leaving retailers to weigh brand recognition against shelf economics.
More
- The article distinguishes Huiyuan Group, which announced it would shift to the new Zanjia Guoyuan brand, from Beijing Huiyuan, which holds the registered trademark and continues selling Huiyuan juice.
- A court ruling cited by the article found specific affiliated companies liable in a trademark and unfair-competition case; the report cautions against extending that ruling to every related entity.
- In stores visited in Jinan, the established Huiyuan products were still available while the new label was not found, illustrating the practical challenge of replacing a brand before the old one disappears.
- Reported shelf prices for comparable two-carton products were higher for the new brand, but the article notes that retail prices alone do not reveal channel margins or consumer demand.
- The comparison with the 2012 Wanglaoji-JDB split highlights that a new name needs supply, distribution and consumer recognition, not just access to the old company's production resources.
Changan shifted its listed-company leadership as sales fell and profit contracted, clarifying responsibilities between the new state-owned parent and the automaker while demanding faster execution.
More
- Zhao Fei became chair of the listed automaker, while Ye Pei took the president role and former chair Zhu Huarong moved to focus on the parent China Changan group.
- The company reported nine-month sales down 18.1% year over year and first-half net profit down 64.3%, according to the article.
- The restructuring separates group-level strategy and resource allocation from the listed company's day-to-day operating responsibility, while Zhao's dual role preserves a direct link between the two.
- Changan says it wants to shorten decision cycles despite the added layer of a newly established state-owned group; the proposed process changes still need to show up in results.
- Management is shifting growth ambitions toward overseas markets, but the article notes that local product design, manufacturing and service capabilities will matter more than simply exporting domestic models.
Chinese coverage presents the proposed AI export-control catalog as boundary-setting, not a broad retreat from openness, noting that DeepSeek and MiniMax completed registration for open releases and overseas licensing.
More
- The Ministry of Commerce and Ministry of Science and Technology sought public comment on revisions to a catalog covering restricted and prohibited technology exports.
- The proposal includes 23 AI-related areas, including large models, agents, computing clusters and information processing.
- The article says DeepSeek and MiniMax completed registration procedures and then carried out open-source releases and overseas licensing.
- Chinese official messaging, as presented here, spends more time explaining what the rules do not regulate than spelling out restrictions, framing the move as delineation rather than closure.
- The practical boundary is not yet clear from this short account; the final catalog and how authorities apply it will determine whether firms face meaningful constraints on international distribution.
Chinese coverage contrasts Duet’s professional focus and approval gates with more consumer-oriented agents, making restraint and permission boundaries central to Anthropic’s product pitch.
More
- Duet is described as a proactive assistant that summarizes priorities and can prepare work across connected services.
- Actions such as sending messages, paying, signing documents or deleting data require user approval; some financial actions are excluded entirely.
- The product is positioned for professionals and small businesses rather than casual household tasks, differentiating it from mass-market assistants.
- The design reflects a hard product trade-off: background monitoring and tool access can save time, but excessive autonomy risks user trust and costly mistakes.
- The report is based on launch materials and selected demonstrations, not an independent evaluation of reliability.
A 908-person survey says respondents born in the 2000s saved an average of 25,400 yuan last year (roughly $3,600), with many buying insurance to cover health and income shocks.
More
- The survey reports illness and job loss as leading financial worries, each cited by roughly six in ten respondents.
- Respondents favored liquid savings: 55.1 percent used flexible accounts, while 43.6 percent kept money in bank accounts and 39.2 percent used time deposits.
- About 64.9 percent said they had commercial insurance through online channels; the survey estimates average cumulative coverage of about 1.34 million yuan (roughly $190,000).
- The pattern suggests young consumers are building buffers against income volatility rather than saving only for major purchases; the sample is modest and survey-based.
Swedish legal-AI firm Legora says it will expand its agent platform; the article notes that rival Harvey trained a legal model on China's Kimi K3.
More
- Legora says its annual recurring revenue surpassed $200 million, with its customer base spanning law firms and in-house legal teams.
- The China-relevant detail is that Harvey's legal model was post-trained from Kimi K3, showing a Chinese model contributing to a foreign specialist-AI product.
- The article presents a crowded market in which workflow integration, source-backed outputs, and lawyer review matter more than generic chat.
A new index aims to compare computing services across hardware and providers, using a standardized unit intended to make heterogeneous capacity easier to price.
More
- The platform is intended to gather supply, demand, transaction, and delivery data to support comparisons across compute types and regions.
- Its proposed standard unit, called one unit of computing power, is meant to convert different hardware and cluster configurations into a comparable service measure.
- The underlying policy ambition is to make compute legible to companies, investors, and local planners, potentially helping shift from headline capacity to useful output.
- A private vendor is helping define the measurement approach, so governance, data coverage, and independent adoption will determine whether the index becomes a genuine market benchmark.
Lilly's deal with China's Jitai centers on continued AI-assisted lipid nanoparticle discovery, positioning a reusable delivery platform as a research partner rather than licensing a single drug candidate.
More
- Jitai receives an upfront payment in the tens of millions of dollars, with up to $1.1 billion in development, regulatory and commercial milestones plus tiered royalties, according to the article.
- The collaboration focuses on lipid nanoparticle delivery technology, with Jitai continuing to generate and test new carriers through its NanoForge platform.
- That structure differs from conventional license-out deals centered on one defined molecule: Lilly is paying for access to an ongoing pipeline of discovery options.
- LNP delivery is a meaningful bottleneck for nucleic-acid medicines because carriers must reach target tissues, enter cells and release their payload; liver targeting is better established than many extrahepatic targets.
- The article says generative design can explore large virtual chemical spaces and prioritize candidates for wet-lab testing, but the deal's value will depend on experimental results and milestones, not the platform label alone.
BroadLink's CEO argues that small specialist models, not general-purpose chatbots, fit cost-sensitive appliances; the company says its system can cost one yuan per device per year.
More
- The supplier's case rests on three constraints for appliance AI: predictable decisions, low latency, and very low operating cost.
- Its proposed architecture uses a set of small specialist models, with a larger model acting as a reviewer that analyzes errors and helps improve them.
- The company says its system costs about one yuan per device per year (roughly $0.14), though the article offers no independent cost breakdown.
- BroadLink sees OTA updates to existing connected appliances as a more practical route than redesigning hardware, given certification and supply-chain costs.
Renault's China R&D center is taking on global vehicle programs and spreading faster development practices across the group, as European automakers look to China's supply base and engineering cadence.
More
- The Advanced China Development Center has about 200 people, compared with roughly 5,000 in Renault's French engineering organization, yet it has worked on the Twingo family and the new Dacia Spring.
- Renault says the center combines Chinese development speed and supplier access with the group's existing engineering and quality processes; the approach is not simply a transfer of work to a lower-cost site.
- The Twingo program is presented as evidence that parallel cross-functional development can compress schedules, with teams responding to changes together rather than passing work sequentially.
- Renault also wants the China team to strengthen intelligent-cockpit and smart-chassis capabilities, areas where the company sees a stronger local ecosystem.
- The next test is whether multiple 2028-era electric models can deliver competitive cost and quality at scale; a single successful program does not establish a group-wide transformation.
Chinese coverage links a sharp September car-sales decline to expensive, scarce memory chips, showing how AI infrastructure demand may be spilling into the cost base and launch timing of another industry.
More
- Passenger-car retail sales in China fell 28% year over year in September, while exports rose 73.7%; exports helped support wholesale volumes despite weak domestic demand.
- The article says new-energy vehicle penetration reached 67% even as those sales declined, while gasoline-car retail fell 43%, illustrating a shift in mix beneath the overall downturn.
- It attributes slower new-model launches and higher vehicle costs partly to memory-chip scarcity as AI demand competes for semiconductor capacity.
- Automotive margins are thin, so even a few thousand yuan of additional memory cost per vehicle can matter to manufacturers; the article uses Xiaomi as an example of reported pressure on margins.
- The story connects two supply chains, but the causal weight of memory prices versus subsidies, the prior-year sales base and consumer demand remains uncertain.
Volcano Engine is testing its video-generation model with 20 studios and directors, using low-resolution previews before selected shots are regenerated in 4K.
More
- The limited rollout is tied to specific productions rather than a broad public launch, with the company exploring whether generated footage can meet premium film and television needs.
- A studio told the outlet that previewing in 480p before generating selected shots in 4K makes the workflow more affordable than expected.
- The strategic test is whether consistency and performance control improve enough for high-end production, not simply whether the model can produce sharper images.
- The article says some participating projects may reach broadcast or film audiences soon, but provides no independent quality evaluation.