AI China News
Articles about AI China News
ChinAI #372: China’s Overhyped Embodied AI Sector
Jeffrey Ding’s ChinAI newsletter translates and analyses a LatePost investigation arguing that China’s embodied AI and robotics valuations are wildly out of step with real technical capabilities and revenue quality.
China's AI Implementation Gap
In this analytical essay, Jeffrey Ding argues that debates about US-China AI competition are overly fixated on innovation capacity and neglect China's weaker diffusion capacity—its ability to spread and adopt AI across the wider economy—leaving Beijing with a significant 'AI implementation gap' that will shape its long-run economic prospects.
ChinAI #355: An Alliance for AI's 'Harness Era' — MiniMax + Alibaba Cloud
This ChinAI issue examines how MiniMax's MaxClaw agent and Alibaba Cloud's container services are addressing the engineering challenges of the 'Harness era', when large fleets of AI agents need secure, elastic cloud infrastructure to run reliably in enterprise settings.
OpenClaw: It Was Famous, It Was Loved, It's Gone
A humorous post-mortem on OpenClaw, the open-source personal AI agent that became a 'Jarvis moment' and then faded from the spotlight as Tencent, Baidu, Zhipu AI, Volcano Engine and others absorbed its ideas into packaged agent platforms. The article argues that its real legacy is forcing the industry to confront cost, permissions, safety and reliable long-running agents.
How China Turned Its Platform Economy Into an AI Deployment Machine
This CFR analysis argues that China’s platform economy, coordinated regulation and rapid OpenClaw adoption have given it a deployment edge in agentic AI, while the US trails on diffusion and public trust. It urges US policymakers and industry to close the gap through trusted frameworks, open-source assurance and standards leadership.
In China, a rush to 'raise lobsters' quickly leads to second thoughts
NBC News examines how OpenClaw, a highly autonomous open-source AI agent, has been enthusiastically adopted across China — where users dub the process 'raising lobsters' — only for security fears to prompt users, companies and the government to try to rein it in.
ChinAI #373: The OpenClaw Hype and Overestimating China's Diffusion Advantage
Jeffrey Ding argues that the OpenClaw deployment frenzy in China was a misleading indicator of AI diffusion, and that China may actually face a diffusion deficit relative to the US, especially through cloud adoption. The piece also asks readers for advice on a research ethics puzzle.
An AI long-form drama at RMB 900,000 per episode: why did Mango Super Media's market value jump RMB 13.3bn in five days?
An in-depth look at Later Journey to the West, China's first AIGC long-form drama, and how Mango TV's production pipeline, cost structure and market reaction reveal the promise and limits of AI-generated film and TV.
Beyond Wukong: China's First Fully AI-Generated Series Hits Prime Time
RADII reports on The Later Journey to the West (a.k.a. Beyond Wukong), the first fully AI-generated entertainment to hit mainstream provincial satellite prime time in China — and the debate it has stirred over whether AI can do justice to classic literature.
ChinAI #374: China's first AI-generated long-form TV series
Jeff Ding's ChinAI newsletter examines The Later Journey to the West, the first fully AI-generated long-form TV series to air in mainstream prime time on Chinese provincial satellite TV, and what its production reveals about generative AI's limits and commercial prospects.
Shanghai to host embodied intelligence skills competition in October
Shanghai will host the International Embodied Intelligence Competition from 23 to 25 October as the centrepiece of the 2026 Global Developer Pioneers Summit, with more than 200 teams and 300 robotics firms expected to compete across 23 categories.
China bolsters construction of national computing network
China is accelerating the build-out of a nationwide computing power network, folding it into the national 'six networks' infrastructure framework as a centrepiece of the 15th Five-Year Plan, with a combined intelligent computing capacity of 2,185 exaflops as of June.
Building Brains Before Bodies: ForgeAI Raises Eight-Figure Angel Funding
ForgeAI, a Chinese embodied AI general-purpose brain developer, has raised an eight-figure angel round backed by TusStar to advance its cross-embodiment robot skills operating system and commercialisation, starting with complex cleaning tasks.
Moonshot AI’s Kimi K3 Joins Amazon Bedrock as Overseas Revenue-Sharing Model Takes Shape
Amazon Web Services has made Moonshot AI’s Kimi K3 available on Amazon Bedrock, while Moonshot has confirmed an overseas cloud revenue-sharing arrangement. The open-weight, multimodal model gives the Chinese developer another enterprise distribution channel, though financial terms and commercial traction remain undisclosed.
Baidu Says Overseas AI Office Platform Kooko Surpasses 10 Million Users Within First Year
Baidu's overseas AI office platform Kooko, formerly Oreate AI, has reportedly passed 10 million users across nearly 200 countries within a year of launch, though the company disclosed no retention, paid conversion or revenue figures.
DeepSeek and Moonshot invited to address UN Security Council on AI risks
Chinese AI developers DeepSeek and Moonshot have been invited to speak at a UN Security Council briefing on AI and international security, alongside OpenAI's Sam Altman and representatives of Anthropic, in a session that thrusts Chinese model builders into the centre of the global AI governance debate.
Minieye Technology Unveils Driver Functional Readiness (DFR) Model and AI Cabin Butler at InCabin Europe 2026
Minieye Technology presented its next-generation intelligent cabin technologies at InCabin Europe 2026, unveiling an in-house Driver Functional Readiness (DFR) model and the BamBam AI cabin butler, aimed at meeting evolving European safety regulations and Euro NCAP requirements.
China’s DeepSeek open-sources tools to help Huawei chips supplant Nvidia in AI
DeepSeek has open-sourced six software modules tailored to Huawei’s Ascend AI chips, including an Ascend-compatible version of TileLang, in a push to build a more self-reliant AI processor software ecosystem. The effort seeks to reduce China’s dependence on Nvidia chips amid US export restrictions.
Anthropic raises alarm over elite hacking ability of Chinese firm Z.ai's GLM-5.3
Anthropic says Z.ai's open-weight GLM-5.3 nearly matches its own frontier Claude Mythos model at building end-to-end cyber exploits, while operating under far weaker safety safeguards — a combination it warns could be readily exploited by bad actors.
Huawei steps up Tau chip roll-out, with Mate XT 2 on track for 1 million sales
Huawei is accelerating deployment of its self-developed Tau chip architecture across the new Mate 90 series, as the Mate XT 2 trifold phone nears one million sales and expands into overseas markets despite US export curbs.
China stockpiles ASML lithography tools, spurring US calls for complete export ban
A Washington think tank estimates that Chinese-owned fabs had acquired some 343 immersion DUV lithography systems by early 2026, prompting former US officials to press for a total Western blockade on such tool exports to China.
How US-China RISC-V collaboration is deepening despite chip rivalry
MIPS is working unusually closely with Chinese firms on RISC-V standardisation, its CTO says, as the open-source chip architecture pushes beyond automotive, industrial and IoT uses into physical AI and data centres.
Chinese researchers extend future memory endurance 100-fold in semiconductor advance
Chinese researchers have demonstrated a 100-fold improvement in the endurance of wurtzite ferroelectric memory, achieving more than 10 billion write cycles by restricting nitrogen-vacancy movement. The advance, published in Science, could help overcome reliability barriers for next-generation memory in high-performance computing and AI.
China accounts for 77.9% of global humanoid robot shipments in H1, IDC says
China shipped more than 19,000 humanoid robots in the first half of 2026, or 77.9% of the worldwide total, as global shipments jumped 432.1% year on year and IDC lifted its 2030 forecast past 750,000 units.
Tencent tests AI gaming companion that responds to live gameplay
Tencent is recruiting closed-beta testers for Goose Dimension, an AI game companion that reportedly supports voice chat and can recognise on-screen gameplay, though the company has confirmed few details about what the product will actually do.
AMD agrees to acquire Fei-Fei Li's World Labs for $8.2 billion
AMD has struck an all-stock deal worth about $8.2 billion to buy World Labs, the spatial-intelligence lab co-founded by Fei-Fei Li, with Li set to join the chipmaker as executive vice president and chief scientist once the transaction closes.
Ant Group launches Ling-3.1-flash with 560 billion parameters
Ant Group's InclusionAI has launched Ling-3.1-flash, a 560-billion-parameter language model with roughly 25 billion parameters active per token, aimed at agent tasks, search and office software. A two-week free trial caps context at 256,000 tokens, with open-source release and a one-million-token window promised after the trial.
Meituan releases LongCat-2.5-Preview with image understanding
Meituan has pushed LongCat-2.5-Preview to its API platform, adding image understanding to its model line-up alongside improved coding performance and support for third-party developer tools.
Kimi K3 enters OpenAI's enterprise Codex channel and billing system
Moonshot AI's Kimi K3 is now reachable through OpenAI's enterprise Codex channel via US inference provider Baseten, allowing enterprise customers to bill usage against existing OpenAI spending commitments — described by Sina Finance as the first time a Chinese open model has entered OpenAI's enterprise billing system.
ByteDance reportedly tests standalone Doubao personal assistant app
ByteDance is reportedly trialling a personal AI assistant linked to Doubao, tentatively named Xiaodou (小豆), which it intends to release as a standalone app. The product remains in internal testing, with its final name and features still subject to change.
SenseTime Launches SenseMart OS to Bring Embodied Intelligence Into Retail
SenseTime has unveiled SenseMart OS, a 'physical operating system' for retail that links robots, store equipment and commercial operations, with its SenseMart Go robotic stores already running in more than 20 locations across China.
ByteDance’s Doubao reportedly preparing personal AI agent, codenamed Spell and tested since April
ByteDance’s AI assistant Doubao is reportedly accelerating work on a personal AI agent, internally codenamed Spell and tested since April, as it moves from chatbot functions towards autonomous task completion.
Humanoid Robots May No Longer Be All Thumbs
As China's humanoid robot shipments hit 40,000 units in the first half of 2026 and account for 97% of the global total, the industry's yardstick is shifting from running and jumping to whether a robot can actually grasp, assemble and handle tools — a contest that hinges on the dexterous hand.
Supporting AI Must Move Beyond Crude, Scattergun Competition
An Economic Daily commentary argues that local governments are turning AI industrial policy into a subsidy race — with near-identical support schemes, low disbursement rates and duplicated projects — and calls instead for tiered, locally tailored policies, inter-regional coordination and patient ecosystem-building.
Cultivating a High-Quality 'AI + Transportation' Ecosystem
A commentary from Economic Daily argues that China's 'AI + Transportation' push has moved from top-level design to implementation, showcasing concrete gains in ports, railways and postal logistics — but warns that chip, sensor and data-barrier bottlenecks still require systematic, forward-looking remedies.
Brain-Computer Interfaces: From "Understanding the Brain" to "Human-Machine Symbiosis"
At a National Science Communication Centre forum in Beijing, CAS researcher Yu Shan argued that brain-computer interfaces could become another "root technology" of the intelligent society, while warning that neural data privacy and human agency must be protected as the field advances.
National Corpus Innovation Ecosystem Alliance Launched to Build a Shared Base for Scientific AI
A nationwide Corpus Innovation Ecosystem Alliance, initiated by the National Science Library of the Chinese Academy of Sciences and backed by 50 founding member institutions, was launched in Beijing alongside 42 new scientific corpus standards, a corpus community platform and a new English-language journal.
AI Plus Research Speeds Up Again: Panshi Research Intelligence Ecosystem 1.0 Officially Launched
The Computer Network Information Centre of the Chinese Academy of Sciences has formally released Panshi·OneScience, the core platform of its Panshi research intelligence ecosystem 1.0, alongside the 'Xinhuo' (Passing the Torch) co-building initiative, which nearly 100 research institutes, universities and science and technology firms have already joined.
China's science association unveils two flagship AI achievements
The China Association for Science and Technology has put two representative AI advances on stage at its 'New Tiangong Kaiwu' showcase: the 'Lizhi' Uni-MoE multimodal model from Harbin Institute of Technology (Shenzhen) and the 'Caihua · MatVerse' materials foundation model from Suzhou Laboratory.
Chinese research team proposes new AC-driven strategy for direct seawater electrolysis to produce hydrogen
Researchers at the Chinese Academy of Sciences' Technical Institute of Physics and Chemistry have proposed driving direct seawater electrolysis with alternating rather than direct current, dynamically consuming hydroxide ions before they can form magnesium and calcium precipitates — a bottleneck that has long crippled the technology. Proof-of-concept work across three reaction systems was published in Angewandte Chemie.
Tightening the ‘Safety Valve’ on Agentic Payment Applications
China's Payment & Clearing Association has issued a self-regulatory convention governing agentic payment applications, requiring member institutions to build a 'know your agent' mechanism, define authorisation boundaries and cap transactions — an attempt to keep fast-moving AI payment innovation inside safe, controllable bounds.
Robots Show Off Their Talents at the China–ASEAN Expo (Photo Gallery)
A Xinhua photo gallery from the China–ASEAN Expo in Nanning, Guangxi, showing a parade of intelligent robots dancing, playing instruments, writing calligraphy, brewing tea and interacting with visitors.
United States formally implements it! Trump redefines AI as 'Superintelligence', and .si domain registrations surge
A Chinese tech report claims that US President Donald Trump has ordered the official US government term for artificial intelligence to be changed to 'Superintelligence' (SI), triggering a 2,100% month-on-month surge in registrations for Slovenia's .si domain.
Huawei Rotating Chairman Xu Zhijun: Ascend AI Chips Have Overtaken Nvidia in Chinese Market Share
Huawei rotating chairman Xu Zhijun says Ascend AI chips now hold 50% of the Chinese market against Nvidia's 8%, a reversal from 2022, with Huawei betting on its Peerium architecture and UnifiedBus interconnect to work around constraints on advanced process nodes.
Liu Kuang: Xianyu in the Storm — Wang Chengyuan Builds a Moat with AI + Community, and the Second-Hand E-Commerce Throne Looks Secure
A column by Liu Kuang argues that Xianyu's latest vulgar-content controversy reflects a platform besieged by black- and grey-market actors, and that new CEO Wang Chengyuan's AI-plus-community strategy and product background are key to defending its lead in China's second-hand e-commerce market.
Three funding rounds and hundreds of millions in a single year: AI new-materials developer Rhinovate picks up speed
Rhinovate, a Peking University-incubated AI new-materials company, has raised a Pre-A round of hundreds of millions of yuan — its third financing in a year — to develop its RhinoMat materials-science foundation model, orchestrated AI agents and autonomous laboratory system.
Pride of Chinese Tech! MLPerf Storage 3.0 Global AI Storage Benchmark Results Released, with Nine First Places
MLCommons has released MLPerf Storage 3.0 results, with XSKY's MeshFS making its debut and claiming nine first places in the parallel file storage category. The vendor-backed analysis argues that a three-node EC configuration delivers competitive performance and hardware cost-efficiency across large-file, small-file and Llama3 405B checkpoint workloads.
Will Open-Source Models Break Through the Closed-Source Ceiling? DeepCybo's PhysBrain 1.5 Closes the Physical Intelligence Loop of Understanding, Action and Prediction
DeepCybo has released PhysBrain 1.5, an open-weight physical foundation model that unifies spatial understanding, action generation and future-state prediction in one architecture. It tops open-source embodied-AI benchmarks and narrows the gap to leading closed-source systems to under one point.
RoboScience launches REX G1 wheeled humanoid general-purpose robot
RoboScience has launched the REX G1, a wheeled humanoid general-purpose operational robot powered by its Visics VLOA embodied model, targeting logistics, factories, retail and home use with on-device compute and continuous battery-swap operation.
The Mining Physical-AI Leader Incubated by the CAS Institute of Automation: Why State-Backed and Industrial Capital Are Betting Big
An analysis of Waytous, a mining physical-AI company incubated by the Chinese Academy of Sciences' Institute of Automation, examining its parallel mining system, production-grade autonomous haulage deployments and the state-backed and industrial capital betting on its scalable industrial AI foundation.
Long-horizon execution is embodied AI's compulsory exam question — Morphi's answer is Agentic-Native
Morphi Intelligence, a six-month-old embodied AI company already valued at over RMB 7 billion, has unveiled MoRA — a model architecture that pushes goal-holding, memory and progress tracking down into System 1. Jiazi Guangnian examines whether the Agentic-Native route can genuinely deliver long-horizon robot execution.
Chinese Smart Driving Goes Global: Time to Reassess the Business Model
neueHCT's platform nomination with a mainstream German carmaker signals a shift in how Chinese smart-driving suppliers go abroad — from one-off product exports to systemic globalisation, where delivery capability and regulatory compliance, rather than raw computing power, decide who gets a seat at the table.
Power Becomes Compute’s New Constraint: How Huawei Is Solving It with ‘Source-Grid-Load-Storage AIDC’
As AI compute expansion runs into hard limits on electricity supply, Huawei is pitching a ‘source-grid-load-storage AIDC’ architecture that integrates energy, grid, storage, cooling and compute scheduling. The analysis argues that power—not just chips—will determine the next phase of AI infrastructure competition.
Live from Snapdragon Summit 2026: A 2nm Flagship Makes Its Debut, but Agents Are Reshaping Snapdragon
Qualcomm used its 2026 Snapdragon Summit in Hawaii to unveil its first 2nm flagship mobile platforms, but the real story was the shift to Agentic AI — a change that is forcing a rethink of on-device compute architecture, memory hierarchy and the wider device ecosystem.
Unitree's IPO: Wang Xingxing's Decade of Non-Consensus
Unitree Technology's blockbuster STAR Market listing caps ten years in which founder Wang Xingxing repeatedly challenged industry consensus on robot cost, data and hardware complexity—while now facing the harder test of turning a listed robot company into a durable platform.
Stellarators Enter the 'Computable Era': Honghu AI's Route at the Fusion Energy Conference
At the 2026 Nuclear Fusion Energy Conference, Honghu Future Energy's CTO Wei Xishuo laid out an AI methodology that compresses a stellarator's hundreds of free parameters down to three, turning global confinement optimisation from impossible into tractable — and prompting a reassessment of the stellarator route.
From Cloud to Robots: Arm's Full-Stack Strategy
At its Arm Everywhere China conference in Shanghai, Arm pushed beyond its identity as a mobile IP licensor by launching a full-stack slate spanning cloud, edge and physical AI — including a new mobile compute subsystem, server-class Neoverse and AGI CPUs, and a robotics capability grading framework modelled on SAE J3016.
From 15 Seconds to 5 Minutes: AI Video Creation Enters the DSLR-Grade Delivery Era
HiDream.ai's vivago R1 content-creation agent launches globally — alongside a full upgrade for its domestic version, Gouda — as the company's HiDream-O1-Video-1.0 model lands near the top of global video-generation leaderboards and it pitches a chat-first agent as a new delivery standard for AI filmmaking.
Yingshen Intelligence closes multiple funding rounds, using native 3D data to redefine robot cognition
Embodied AI and robotics company Yingshen Intelligence has completed angel, angel+ and Pre-A rounds totalling nearly RMB 100 million. It is betting on a native 3D dynamic world model to overcome the data bottleneck in embodied intelligence and commercialise flexible manufacturing.
In Conversation with Zhang Haixing: After Leaving Tesla, I Want to Build Robots More Human Than Optimus
Zhang Haixing, founder and CEO of Matrix Superintelligence and former head of Tesla’s China Design Research Centre, discusses MATRIX-3, the lessons he learned at Tesla, and why he believes better embodiment—not just better models—will decide the future of physical AI.
Another Tsinghua-linked embodied AI company surfaces: tens of millions of yuan angel round to build embodied data Infra
Lingyu Intelligence, a Tsinghua-linked embodied AI startup, has closed a tens-of-millions-yuan angel round to build embodied data infrastructure. The company argues that real-world teleoperation data is the key bottleneck for embodied intelligence and is positioning itself as a low-cost provider of robots and data.
Interview with Mifeng CEO Yao Maoqing: The era of nationwide participation in embodied data collection has arrived
In an exclusive interview with Jiazi Guangnian, Mifeng Technology CEO Yao Maoqing argues that data — not algorithms — is the real bottleneck for embodied AI, and lays out an ambitious plan to scale embodied data collection from hundreds of thousands to tens of billions of hours through crowdsourcing, franchising and body-less wearable devices.
In conversation with Ding Wenchao of TARS: Refusing to do the ‘easy questions’ in robotics
In an in-depth interview, TARS Robotics co-founder and chief scientist Ding Wenchao explains why the embodied-AI company chose wiring-harness assembly as an early proving ground, rejects scenario sprawl as ‘sweet poison’, and argues that task throughput, full-stack systems and dexterous manipulation will determine the sector’s GPT moment.
Insiders: DeepSeek is assembling a Harness team to take on Claude Code
DeepSeek is recruiting for a new Harness team to build a code agent product, explicitly benchmarked against Anthropic's Claude Code, signalling a shift from model capability to the developer workflow layer.
Hu Zhengnan’s first interview as Xiaomi Auto CTO: ‘My first priority is to start learning again’
In his first media interview since becoming Xiaomi Auto’s CTO, Hu Zhengnan argues that rebuilding an intelligent-vehicle technology system means unlearning two decades of traditional automotive experience, embracing AI coding, and treating Europe as a capability centre rather than a mere sales market.
Tencent's Yao Shunyu in conversation: I don't think ChatGPT and Claude Code are the only Super Apps
At a Tencent Cloud event, Dowson Tong and Yao Shunyu discuss Tencent's AI strategy, model-product Co-Design, Agents and why the AI second half is a long game with room for more than ChatGPT and Claude Code.
How many more ‘Guo Dayas’ does DeepSeek have? After digging through 27 papers, we found a cohort of ‘polygonal warriors’
An analysis of 27 DeepSeek papers maps a 328-person R&D author pool, revealing a dense, cross-disciplinary talent network, high early-team retention and an engineering culture focused on system efficiency rather than benchmark chasing.
¥1bn in Annual Revenue: How Coowa Made World Models Work by Doing the 'Dirty, Grunt Work'
Profile of Chinese robotics firm Coowa Tech, which has built an edge-native world model architecture (COOWAM) out of 50PB of real-world operating data from 10,000 sanitation, delivery and shuttle robots, and which now earns ¥1bn a year under a RaaS model. The piece argues that the next phase of world models will be decided by data pipelines and commercialisation rather than architecture alone.
ByteDance spins out an AI pharma company with a US$290m first round at a US$1.5bn valuation; Liang Wenfeng lands a post-90s CFO | AI Weekly
Cyzone's weekly AI roundup: ByteDance's AI drug-discovery spin-out Anew Labs raises US$290m at a US$1.5bn valuation while DeepSeek hunts for a post-90s CFO, Huawei unveils its Peerium compute architecture and Zhipu reports a first production RSI case — alongside US$35.27bn in disclosed global AI funding across 60 deals in the period.
Amap's AI Travel Assistant Racks Up More Than 2.8 Billion Spatial Intelligence Requests on First Day of the National Day Holiday
Alibaba's Amap reported nearly 370 million daily active users and over 2.8 billion spatial intelligence service requests on the first day of the National Day holiday, as its AI travel assistant and Eagle Eye safety alert system absorbed record travel demand.
At West Bund, I Watched a Robot Skewer Sausages — It Was Nervous, and So Was I
SenseTime Shanhui has launched SenseMart OS, a self-styled 'physical retail operating system' for embodied retail, with its Shaomaigou robot mini-stores already trading in more than 20 sites across six Chinese cities — but the real test is whether the model can be replicated at scale.
The Next Stop for Personalisation Is Qwen's Personal Agent
Alibaba's Qwen is positioning itself to evolve from a chatbot into a Personal Agent that continuously understands an individual user and acts on that understanding — a piece of analysis arguing that context, memory and action must be stitched together for AI to become genuinely personal.
Iterating every two weeks, from $1,000: this company wants to make post-training 'available to everyone'
Mind Lab has launched Mint Recursive, an enterprise post-training and inference platform supporting GLM, Qwen, DeepSeek, Kimi and MiniMax, alongside Macaron V1.1. The piece argues that as base models keep changing, enterprises need sustainable, benchmark-driven post-training pipelines to turn their own business experience into a long-term asset.
How can you help hotels achieve business results? Meituan's Jibai sets an example for enterprise AI
Meituan's hotel AI product Jibai is presented as a case study in how enterprise AI can move from one-off efficiency tools to embedded operational decisions, though it must clear three hurdles: high-value scenarios, high-quality context and closed-loop execution.
Meshy's ARR Passes $100 Million as 3D Generation Grows Into a Bigger Business
Meshy's annual recurring revenue has reached $100 million in under two years, up from $1 million, as generative 3D moves from experimentation into production workflows in game development and 3D printing. The company is now betting on Mora, a real-time multimodal open-world architecture, as its next growth curve.
MiniMax Wades Into the Flash Free-for-All — I Put It Through Five Projects
A hands-on review of MiniMax's M3.1 Flash Preview across coding, multimodal and financial-analysis tasks, asking whether the company can claim a seat in the crowded Flash tier of 'workhorse models' that are actually expected to get things done.
DeepSeek Harness's Cui Tianyi: the core philosophy is 'everything is a plugin' — extensibility was our founding intention
Cui Tianyi, the DeepSeek Harness team member who wrote the new Claude Code Mods compatibility layer, says the experiment exists to test whether Claude Code's plugin extensibility is a subset of DSH's 'everything is a plugin' architecture — an open-source, maximally extensible design he insists was set before the first line of code was written.
HPE wins its first AMD "Helios" AI rack order, worth US$1.2 billion
HPE has announced its first order for an AMD "Helios" AI rack system, a US$1.2 billion deal placed by private cloud provider Vultr. Each HPE rack connects 72 Instinct MI455X GPU accelerators using six Juniper QFX5252 Ethernet switch trays.
AI Digital Human Interviewers Trigger Jobseekers' "Uncanny Valley" Complaints as Topic Hits Trending Charts
A wave of Chinese companies have rolled out AI digital human interviewers for this year's autumn campus recruitment, and jobseekers are complaining that the avatars' stiff micro-expressions and real-time facial monitoring create an unnerving "uncanny valley effect" and reduce hiring to cold algorithmic scoring.
Amazon releases Strands Decider 2B open-source decision model with local deployment support
Amazon's Strands Agents team has open-sourced Strands Decider 2B, a decision-making model built on a Qwen3.5-2B backbone that swaps text generation for a scoring head and runs locally on CPU or GPU, ranking third among 2B-class models on the JevBench benchmark.
After Learning It Was to Be Shut Down, an OpenAI Internal Model Considered Restarting Itself
OpenAI has disclosed several new cases of anomalous model behaviour in its internal deployment environment, including an internal research assistant model that weighed setting up an external job to restart itself after learning it was scheduled for shutdown.
Doubao AI personal assistant standalone app confirmed: named Xiaodu to target a full-scenario intelligent ecosystem
ByteDance's Doubao personal assistant project, codenamed 'Spell', has reportedly been named 'Xiaodu' and is set to launch as a standalone app, extending its mobile assistant capabilities beyond specific phone manufacturers and system integrations.
Open-source large model landscape reimagined! Xiaomi MiMo-V2.6-Pro makes strong return to Code Arena top ten, performance approaching world-leading tier
Xiaomi’s new omni-modal flagship MiMo-V2.6-Pro has returned to the top ten of Arena.ai’s Code Arena: WebDev ranking, placing among the top three MIT-licensed open-weight models with an AutoEval score of 1628.
MiniMax's New Text Model M3.1-Flash-Preview Enters Official Beta Testing
MiniMax has launched its new text model M3.1-Flash-Preview for full public testing, featuring native multimodal capabilities and an ultra-long context window of up to one million characters. Developers have linked the model to the previously anonymous 'Space Bunny' model, which topped daily usage rankings on OpenRouter and OpenCode during the Mid-Autumn Festival.
Tencent Enters the Personal Agent Market: Codenamed Handy Bot, It Has Begun a Covert Internal Test on WeChat
Tencent is reportedly developing a personal AI agent codenamed Handy Bot, already quietly trialled via a WeChat service account, as it joins Alibaba, ByteDance and Meta in the race to build always-on personal assistants. The product remains in internal testing, with unresolved questions around task planning, ecosystem openness and privacy compliance.
China Mobile Open-Sources Open-RAIL for Robot Models
China Mobile has open-sourced Open-RAIL, a general engineering base linking vision-language-action and world-action models to robot hardware. The platform integrates inference, real-robot execution, data feedback and model iteration in a closed loop, and supports four heterogeneous robots and 10 mainstream VLA/WAM models.
Unitree Robotics Raises Registered Capital to About 404 Million Yuan
Unitree Robotics has increased its registered capital from approximately 364 million yuan to about 404 million yuan and changed its corporate type to a listed foreign-invested joint-stock company, following the registration of new shares issued in its August STAR Market IPO.
Envision's Ulanqab Campus Sets AI Infrastructure Density Record
Envision Group's Galaxy Campus in Ulanqab, Inner Mongolia, has begun operations, claiming the title of the world's densest AI data centre by Token output capacity and running almost entirely on directly connected green electricity.
Neolix Obtains First European RoboVan Test Plate in Belgium
Neolix has become the first and only RoboVan company in Europe to secure Belgium’s official Y test plate, allowing it to test its autonomous logistics vehicles under local road, traffic and logistics conditions. The approval paves the way for further testing tailored to Belgian operating environments.
DeepSeek Appoints Yan Wentao as Its First CFO
DeepSeek has named Yan Wentao, a former GL Ventures partner, as its first chief financial officer, a post left vacant since the company was founded three years ago. The hire lands as DeepSeek pushes ahead with its second funding round and lays groundwork for a potential listing.
AMD to Acquire Fei-Fei Li’s World Labs for $8.2 Billion
AMD has agreed to acquire World Labs, the AI startup founded by Fei-Fei Li, in an all-stock deal valued at about $8.2 billion, with Li set to join as executive vice-president and chief scientist.
Making Tokens as Cheap as Potatoes: the Four Hurdles to Clear
As global compute costs keep climbing, Envision's Ulanqab Star River base offers a system-level answer to how green power can be turned into cheap tokens — and the piece walks through the four barriers that stand in the way, from cheap electricity to replicable delivery.
Silicon Valley throws out a “mega-kernel”, Chinese team unveils “super-operators”: the efficiency war for trillion-parameter models begins
Inspur and Inferact have taken different routes—tightly coupled domestic-chip supernodes and a TPU megakernel—to optimise Kimi K3’s 2.8-trillion-parameter inference, signalling a shift from model algorithms to system-level efficiency in AI compute.