At its chip technology briefing, Xiaomi unveiled the Xuanjie O3 flagship mobile system-on-chip, touting an AnTuTu benchmark of 5.22 million points — the first flagship SoC to clear the 5-million mark and a product the company says took 459 days to develop.
Built as an “AI flagship” SoC for its premium devices, the chip pairs a ten-core, all-big-core CPU (six super-large cores and four large cores reaching 4.35GHz) with a multi-core score that broke 15,000 and delivers a 60 percent performance lift. The GPU is a 16-core G2-Ultra NX with third-generation ray tracing, bumping graphics performance by 85 percent while cutting power draw by 64 percent.
Xuanjie O3 is also the first mobile processor to support LPDDR6 memory, with bandwidth of 113.8 GB/s — a 48 percent gain over Xuanjie O1. The neural processing unit has been rebuilt around Xiaomi’s on-device MiMo models, offering up to 200 TOPS of tensor compute and 3.13 TFLOPS of vector compute, improving on-device AI performance by about 45 percent.
The imaging pipeline upgrades to a fifth-generation flagship ISP that supports a single camera of up to 432 megapixels, three-camera arrays of 64M plus 64M plus 50M, and 24-bit color depth, alongside 4K60FPS AI night-video denoising. A self-developed RISC-V security core now carries both China’s domestic encryption certification and the CCRC EAL5+ assurance grade.
Xuanjie O3 pushes AI across every module: dual SME2 acceleration units in the CPU, eight NX neural accelerators in the GPU, an AINR unit inside the ISP, a hardware AI super-resolution unit in the DPU and an AI engine in the ADSP. Low-power optimizations, including a unified fusion bus and top-layer metal routing, deliver a static latency of 82 nanoseconds.
The release rounds out a three-chip lineup: the O3 now enters mass production, while the O100 and D100 complete their development phase and are set to go commercial next year. The Xiaomi 18 Fold and Xiaomi Pad 9 Pro Max, expected in September, will be the first devices to ship with Xuanjie O3.
","publishedDate","2026-08-24T08:09:18.465Z","featured",false,"viewCount",0,"readingTime",2,"status","published","originalWordPressId","featuredImage","featuredImageUrl","https://cms-image.pandaily.com/1/img_1_8f4700ef9d.png","gallery",[],[],"tags",[344,351,354],{"_316":345,"_11":346,"_319":347,"_348":349,"_34":350},"nx94rpc8tfd6tr1hlo9y0dct","Xiaomi","xiaomi","color","#6366f1","",{"_316":352,"_11":353,"_319":353,"_348":349,"_34":350},"l6utr05w7v1nx4svj5med84i","chip",{"_316":355,"_11":65,"_319":64,"_348":349,"_34":350},"ji5tef72gjtbd7bx0er6q9jv","authors",[],"seo","externalVideo","createdAt","2026-08-24T08:08:39.062Z","updatedAt","2026-08-24T08:09:18.366Z","publishedAt","2026-08-24T08:09:18.424Z","featuredPosts",[368,396,426,454,477],{"_316":369,"_129":370,"_319":371,"_321":372,"_323":373,"_325":374,"_327":375,"_329":330,"_331":376,"_333":334,"_6":7,"_335":-5,"_336":-5,"_337":377,"_56":378,"_342":379,"_356":392,"_358":-5,"_359":-5,"_360":393,"_362":394,"_364":395},"irdlcu5eaoyt70rej8jw63lt","Tencent Hunyuan Releases Hy4 Preview, Ranking Among the Top Tier of Open-Source Models","tencent-hunyuan-hy4-preview-open-source-aug2026","On August 28, Tencent Hunyuan released and open-sourced Hy4 preview, its next-generation large language model with 770 billion total parameters and a context window exceeding 1 million tokens.","Tencent Hunyuan’s Hy4 preview marks the company’s third major model release in six months, reflecting its rapid development cadence and growing strength in foundation models. With 770 billion total parameters, 49 billion activated parameters, and a context window exceeding 1 million tokens, Hy4 preview ranks among the top tier of open-source models and performs strongly across coding, office productivity, and scientific research.
\n\nThe model’s development is closely integrated with Tencent products such as WorkBuddy and CodeBuddy through a model–product co-design approach. In Tencent’s internal blind evaluation, Hy4 preview slightly outperformed GLM 5.3 and Kimi K3, reinforcing WorkBuddy’s position as a leading AI productivity agent in China.
\n\nHunyuan continues to work closely with products such as CodeBuddy and WorkBuddy to optimize the real-world user experience in productivity scenarios. In a blind evaluation conducted internally at Tencent, involving 163 experts and 203 engineering tasks, Hy4 preview achieved an average score of 2.99 out of 4, slightly outperforming GLM 5.3 (2.92/4.00) and Kimi K3 (2.94/4.00).
\n\nThis means that the combined Hunyuan–WorkBuddy experience has achieved a meaningful lead, demonstrating that Tencent AI’s model–product co-design approach has developed into a virtuous cycle.
\n\nPositioned as a model "built for productivity," Hy4 preview has made significant progress across a wide range of real-world productivity tasks. This progress is driven by high-quality data co-developed with Tencent's leading experts in software engineering, gaming, finance, security, and other fields, as well as deep co-design with products such as WorkBuddy.
\n\nIn software engineering, the model has strengthened its ability to understand, plan, debug, and validate long-horizon development tasks, while further improving the visual quality and interactivity of front-end development.
\n\nIn office and analytical scenarios, the model demonstrates substantially better understanding of complex workplace environments and stronger financial analysis capabilities. It places particular emphasis on data analysis and cross-file collaboration, supporting the complete workflow from information processing to the delivery of documents, spreadsheets, and presentations.
\n\nIn game development, the model can generate a playable prototype directly from a one-sentence request and work proficiently with game engines. Developers can continuously refine complex game projects through multi-turn interactions.
\n\nIn scientific research, the model has significantly improved its ability to understand, reason through, and solve complex research problems, making major advances across AI R&D, molecular dynamics simulation, condensed matter physics, foundational mathematics, and other domains.
\n\nNotably, Hy4 preview contributed meaningfully throughout its own end-to-end development process, participating for the first time in the automated optimization of training methods, data strategies, evaluation systems, and low-level operators. The model proposed approaches, ran experiments, and iterated based on the results. Code, logs, and feedback generated by those experiments were then fed into subsequent rounds of exploration, forming an initial closed loop of recursive self-improvement.
\n\nHy4 preview also independently analyzed bottlenecks in the inference system and carried out multiple rounds of optimization in areas such as operator fusion and communication efficiency. These efforts increased end-to-end throughput by 31.8% over the baseline, with consistent gains across different context lengths and concurrency levels—an early demonstration of the model's ability to autonomously optimize inference infrastructure.
\n\nHy4 preview is now open source and has launched simultaneously across Tencent products, including the Chinese and international versions of WorkBuddy and CodeBuddy, Yuanbao, and ima. Users can experience the model directly through these products or access its API through Tencent Cloud TokenHub and OpenRouter. Hy4 preview continues to follow an accessible, cost-effective pricing strategy: $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens for cache hits.
\n\nSince its launch in early March 2026, WorkBuddy has maintained a rapid pace of iteration, continuously improving its user experience and earning recognition from a wide range of individual and enterprise users. According to Tencent’s Q1 2026 earnings report, WorkBuddy has become China’s most popular AI productivity agent service by daily active accounts.
\n\nSince rebuilding its infrastructure in February, Tencent Hunyuan has released a major model iteration approximately every two months. By launching a preview first and following with a full release, Hunyuan continuously incorporates real-world feedback into the development process, allowing the model to improve by solving practical problems. Following this cadence, the next version of Hy4 will begin rolling out in the near future.
\n\nWith its model capabilities now substantially strengthened, Tencent has further consolidated WorkBuddy’s position as China’s answer to Codex, securing an early lead in the rapidly evolving AI landscape.
","2026-08-28T07:32:03.000Z",true,4,"https://cms-image.pandaily.com/1/hy4_preview_f6d614fd20.png",[],[380,384,388],{"_316":381,"_11":382,"_319":383,"_348":349,"_34":-5},"syz0qnsnjbckhahijkip6evy","Tencent Hunyuan","tencent-hunyuan",{"_316":385,"_11":386,"_319":387,"_348":349,"_34":-5},"hjttfsav6py7zihu7gbxd9qc","Hy4","hy4",{"_316":389,"_11":390,"_319":391,"_348":349,"_34":-5},"wbp115vvzeffpp45oh695q2z","Large Language Model","large-language-model",[],"2026-08-28T07:32:03.648Z","2026-08-28T08:07:01.175Z","2026-08-28T08:07:01.243Z",{"_316":397,"_129":398,"_319":399,"_321":400,"_323":401,"_325":402,"_327":375,"_329":330,"_331":332,"_333":334,"_6":7,"_335":-5,"_336":-5,"_337":403,"_56":404,"_342":408,"_356":422,"_358":-5,"_359":-5,"_360":423,"_362":424,"_364":425},"pb5pdb1fuswjlrh8od5anrwd","WorkBuddy Goes to Work at Shanghai Jahwa: 9 Departments, 4.5x Average Efficiency","workbuddy-shanghai-jahwa","Tencent's office agent WorkBuddy has scaled from pilot to production inside century-old cosmetics maker Shanghai Jahwa, lifting efficiency more than fourfold across nine departments.","A century-old Chinese cosmetics maker has become a proving ground for enterprise AI agents. In April, Tencent Smart Retail, together with Shanghai Jahwa's IT and HR teams, launched a pilot of WorkBuddy, the Tencent AI office agent. Months later the deployment has scaled into routine production: WorkBuddy now covers nine business departments — spanning supply chain, brand marketing, R&D, finance, HR and IT — and lifts average employee efficiency by more than 4.5 times.
\nUsage data over the past month shows it is no longer just drafting documents and slides. About 67% of users apply WorkBuddy to specialized vertical work, 63% to data processing and analysis, and 57% to automation and operations. R&D teams use it for cosmetics formulations and regulatory research; finance automates reimbursement-review rules and drafts investment screening reports; HR flags pending probation reviews; supply chain keeps shop-floor dashboards updated daily; marketing taps it for consumer insight and content generation.
\nA supply-chain case illustrates the depth. Jahwa's B2C business spans Tmall, JD and Douyin, with warehouse, handling and delivery costs scattered across seven bill types totaling about 60 million rows over 18 months. The logistics team once wrestled this with spreadsheets, each monthly review taking five to six person-days. Its operations manager, an economics graduate who does not code, broke the process into steps for WorkBuddy to generate Python scripts, validated them against 18 months of history, and packaged what worked into reusable skills. Monthly bill analysis shrank from two person-days to 15 minutes, one visualization to three minutes, and the whole monthly cycle to under 20 minutes — with a single task up to 150 times faster.
\nThe habit is spreading through workplace culture. Employees share skills so a flow a department masters can be reused elsewhere, and one team launched a 70% conversion of weekly PPT reviews into HTML pages that the whole company can view. For Tencent, the case underlines a broader pitch: enterprise agents earn their keep not by replacing jobs, but by letting domain workers — no coding required — hand off the drudgery and turn their own workflows into products.
","2026-08-28T00:00:00.000Z","https://cms-image.pandaily.com/1/A5_c6062cab44.png",[405],{"_316":406,"_11":65,"_319":64,"_348":407,"_34":350},"ou7tkn9cq6eba9jr68hqiwch","#000000",[409,413,417],{"_316":410,"_11":411,"_319":412,"_348":349,"_34":-5},"ppf1i9imlnsp5i6ayv9byk4l","WorkBuddy","work-buddy",{"_316":414,"_11":415,"_319":416,"_348":349,"_34":350},"njwm066t5dgnkyahw9o23jvu","Tencent","tencent",{"_316":418,"_11":419,"_319":420,"_348":349,"_34":421},"ubu5kkvaf94arl01l82x96wt","AIAgent","ai-agent","Artificial intelligence systems designed to perceive their environment, make decisions, and take actions autonomously to achieve specific goals.",[],"2026-08-28T01:48:44.596Z","2026-08-28T01:49:04.868Z","2026-08-28T01:49:04.903Z",{"_316":427,"_129":428,"_319":429,"_321":430,"_323":431,"_325":402,"_327":375,"_329":330,"_331":332,"_333":334,"_6":7,"_335":-5,"_336":-5,"_337":432,"_56":433,"_342":438,"_356":450,"_358":-5,"_359":-5,"_360":451,"_362":452,"_364":453},"tro4f6kqnxy67xhnddulnbs7","Hong Kong Re-Optimizes for Space: From China's Forgotten Launch Pioneer to a New Science Hub","hong-kong-space-science-pioneer","Once China's earliest commercial-satellite pioneer, Hong Kong is burnishing a quieter space story through its universities — even as the gap from science to commerce persists.","Few recaller sees Hong Kong as a space hub, yet in 1990 it stood at the front of China's commercial space ascent. On April 7 of that year, a Long March 3 rocket launched from Xichang carrying the AsiaSat-1 communications satellite into transfer orbit — China's first international commercial launch, executed on behalf of Hong Kong's AsiaSat, the region's first commercial satellite company. Long March rockets soon held about 10% of the global commercial launch market. A second Hong Kong operator, APT Satellite, followed into geostationary orbit. Then the momentum stalled, and the satellite dishes on the Tai Po waterfront became a quiet monument to that early surge.
\nToday the driving force is Hong Kong's universities, whose space science punches well above the city's size. The Hong Kong Polytechnic University has spent decades on lunar and deep-space payloads — sampling mechanisms for the Chang'e-5 and Chang'e-6 missions, camera-pointing systems for Chang'e-3 and -4 and a landing-monitoring camera for the Tianwen-1 Mars probe. The Hong Kong University of Science and Technology sent the MUSICO greenhouse-gas imager to China's space station in 2026, where the first Hong Kong astronaut operates it; it also built the city's first self-developed satellite, Hsiung-bing One. The University of Hong Kong has contributed a lobster-eye X-ray satellite and an instrument for the Chang'e-7 lunar mission, while the Chinese University of Hong Kong has launched two AI remote-sensing satellites that run an LLM on board for real-time image inference.
\nWhat once linked these threads is missing: commercialization. Apart from AsiaSat and APT Satellite, no third sustainable Hong Kong space company has emerged. Decades of de-industrialization, a thin aerospace industrial base and a regulatory bias toward oversight over incentives all held back would-be ventures, including ambitious satellite and constellation plans.
\nHong Kong's advantages are those no other Chinese city can copy: a deep research base, mature international capital markets, common-law commercial rules, global talent and a unique window linking mainland China to the world. Commentators argue hardware manufacturing need not be its answer — research-and-design output, international partnerships and cross-border licensing could be. Whether Hong Kong can convert its forgotten pioneering start into a renewed space economy is now the open question.
","https://cms-image.pandaily.com/1/A4_3188020f29.png",[434],{"_316":435,"_11":436,"_319":437,"_348":407,"_34":-5},"nd7u5jlcbw44fqibmzu4ckhw","Technology","technology",[439,443,446],{"_316":440,"_11":441,"_319":442,"_348":349,"_34":350},"zznmw6zy2dhae9mn0ct3oyr6","Hong Kong","hong-kong",{"_316":444,"_11":445,"_319":445,"_348":349,"_34":350},"u4tbgkcmbtzt3w63dugkjddy","aerospace",{"_316":447,"_11":448,"_319":449,"_348":349,"_34":350},"j45fwu61bg7hx57fo10s37oq","China","china",[],"2026-08-28T01:48:43.659Z","2026-08-28T01:49:04.060Z","2026-08-28T01:49:04.094Z",{"_316":455,"_129":456,"_319":457,"_321":458,"_323":459,"_325":402,"_327":375,"_329":330,"_331":332,"_333":334,"_6":7,"_335":-5,"_336":-5,"_337":460,"_56":461,"_342":463,"_356":473,"_358":-5,"_359":-5,"_360":474,"_362":475,"_364":476},"rt2jkwz0tuqvhpi7fvrg8o8e","Baidu's Dazi Becomes Fastest-Growing Office AI Agent on an 'Application-Driven' ERNIE Bet","baidu-dazi-office-agent","Baidu's office agent Dazi posted the fastest desktop growth among AI work assistants in July as the company doubles down on an application-driven ERNIE strategy.","China's office-AI market is turning into an all-out product war, and Baidu's Dazi has emerged as the fastest riser. According to the AICPB July ranking, Dazi reached 6.74 million monthly active users on desktop, second only to Tencent's WorkBuddy at 11.15 million, but with month-over-month growth of 1,063.79% — the fastest of any office AI agent.
\nBaidu disclosed at its August 27 product release that Dazi users grew nearly ninefold in the past month. Since launching in March, the agent has iterated roughly 150 times, at times shipping nearly a release a day. The enterprise edition now bundles 15 suites and 96 skills spanning finance, legal, product and R&D, operations and HR. Daily questions since launch have grown 60-fold, and website visits rose 845% month over month in July.
\nThe company frames Dazi's edge as organizational rather than purely algorithmic. Executive vice president Shen Dou argued a good agent must not only be a smarter model, but know the user better and marshal tools and professional methods. Dazi leans on Baidu's full-stack AI cloud and its ability to customize delivery, an advantage its competitors—Tencent's communication- and document-leaning WorkBuddy and Alibaba's team-collaboration-focused Qwen—replicate less easily.
\nBehind the product is a broader strategic tilt from chairman Robin Li. At the June AIDAY he proposed 'DAA' (daily active agents) as a new measure of AI value — counting how many agents actually complete tasks every day rather than tokens consumed. In August he spelled out an 'application-driven' course for ERNIE: reinforcement from the real apps that matter most, including AI search, digital humans, Miaoda, Famou and Dazi.
\nThe bet is a self-reinforcing loop: real products tell the model what capability matters, the model optimizes those tasks, adoption rises, and new usage data feeds the next round. As features converge across the industry, Baidu is wagering that whoever runs this application-driven flywheel deepest will be the office AI leader that users actually keep.
","https://cms-image.pandaily.com/1/A3_6c5d54b3a8.png",[462],{"_316":406,"_11":65,"_319":64,"_348":407,"_34":350},[464,468,469],{"_316":465,"_11":466,"_319":467,"_348":349,"_34":350},"w0jddsa493afvf6rc1oeyxxr","Baidu","baidu",{"_316":418,"_11":419,"_319":420,"_348":349,"_34":421},{"_316":470,"_11":471,"_319":472,"_348":349,"_34":350},"am8kdhkyb01i2uiwlomo5wni","office tool","office-tool",[],"2026-08-28T01:48:42.611Z","2026-08-28T01:49:02.956Z","2026-08-28T01:49:02.991Z",{"_316":478,"_129":479,"_319":480,"_321":481,"_323":482,"_325":402,"_327":375,"_329":330,"_331":332,"_333":334,"_6":7,"_335":-5,"_336":-5,"_337":483,"_56":484,"_342":487,"_356":500,"_358":-5,"_359":-5,"_360":501,"_362":502,"_364":503},"k41q9vhv5unni1p3lrbvyshg","XPeng's Second-Gen VLA Adds 'Time,' Pushing Physical AI From 3D Space to 4D Spacetime","xpeng-second-gen-vla-time","XPeng debuts the first major upgrade to its second-generation VLA, adding dynamic 4D spacetime understanding and bringing Robotaxi-grade L4 behavior to production cars.","XPeng has delivered the biggest upgrade yet to its second-generation Vision-Language-Action model, and the core change is deceptively simple: the AI now understands time. At a 'TIME'-themed physical-AI showcase, the company said its foundation model has climbed from static 3D spatial understanding to dynamic 4D spacetime understanding, with the new XOS 6.3.0 software debuting globally on the G9L.
\nThree technologies anchor the leap. Infini-VLA, a long-horizon architecture, remembers up to 30 seconds of road scene, stitching continuous driving events into a coherent temporal understanding. Streaming Inference shifts the model from discrete reasoning to continuous, parallel processing, cutting end-to-end response latency by 300%. X-Foresight, a predictive world model, anticipates the behavior of surrounding traffic participants up to six seconds ahead, sharpening reactions to sudden cut-ins and hard braking. On-device parameters grew 3.5 times — more than 15 times that of mainstream VLA models — and comprehensive multi-dimensional safety improved 20-fold.
\nXPeng also introduced Master Agent, a vehicle-agnostic 'brain' built on its Omni multimodal model that fuses VLA perception with the VLM cockpit, automatically decomposing user intent and coordinating autonomous driving, chassis and cabin systems. New voice-activated features, such as requesting a pull-over by voice, effectively bring Robotaxi-grade L4 capability into mass-production vehicles.
\nThe upgrade arrives as XPeng doubles down on physical AI more broadly. Its robotics business recently closed a first private round of more than $900 million at a valuation above $6.3 billion — a record for a single round in China's embodied-intelligence sector — with backing that includes IDG, Gaorong, Tencent and Alibaba.
\nFor XPeng, the ambition is a unified physical world model that treats driving, robots and cabin intelligence as one continuous, time-aware reasoning problem rather than separate point tasks. By encoding time into the model, it is betting that the next leap in autonomy looks less like sensing and more like understanding the world as it unfolds.
","https://cms-image.pandaily.com/1/A2_bcdb7bfbc3.png",[485],{"_316":486,"_11":73,"_319":72,"_348":407,"_34":350},"sxqv5wtxm67bt4hbdptijger",[488,492,496],{"_316":489,"_11":490,"_319":491,"_348":349,"_34":350},"a27csecjx8sqir4b4lif5rua","XPeng Motors","xpeng-motors",{"_316":493,"_11":494,"_319":495,"_348":349,"_34":-5},"movepp1qjv73p1b86n3pi91s","VLA","vla",{"_316":497,"_11":498,"_319":499,"_348":349,"_34":350},"la9w1kj8q5529bq46ko354xd","Autonomous Driving","autonomous-driving",[],"2026-08-28T01:48:41.457Z","2026-08-28T01:49:02.029Z","2026-08-28T01:49:02.066Z","isDraft"]