Zhipu AI has launched GLM-5.3-Flash, a frontier model it says runs entirely on domestic AI chips, with 100,000 locally made accelerators supporting all of its online inference traffic.

The model reached No. 10 on the AAII leaderboard of the benchmarking firm Artificial Analysis, putting it ahead of DeepSeek's V4 Pro Max. Zhipu AI said the entire workload for GLM-5.3-Flash's online requests is served by 100,000 domestically produced chips.

GLM-5.3-Flash was first released under the codename "Ox Alpha" on August 20, and it topped the usage rankings on the AI model routing platform OpenRouter within a week. The company positions it as a high-performance model that is efficient enough to run at scale on China's domestic compute stack.

Zhipu AI has not publicly identified the chip vendor it used. CNBC reported that analysts suspect the hardware is Huawei's Ascend series. Ivan Lin, an analyst at research firm Counterpoint, said Chinese AI model developers are increasingly directing investment toward AI servers and compute infrastructure built on domestically produced chips.

The launch is the clearest signal yet that at least one of China's leading model labs believes the domestic supply chain is ready to carry flagship workloads. Zhipu AI has been among the most aggressive of China's "AI tigers" in expanding compute, and running a frontier-class model day-to-day on homegrown accelerators is a statement about both performance and chip availability.

The move also fits a broader industry push. As export controls restrict foreign frontier chips, Chinese model builders have raced to optimize their software around the hardware they can actually source. A model that ships at the top of benchmark tables while running on domestic chips is evidence that software efficiency can compensate for hardware gaps.

For Zhipu AI, the Flash line represents a strategy of pairing model quality with practical scalability. With 100,000 chips behind it, GLM-5.3-Flash is designed to show that a China-based lab can deliver competitive frontier performance without depending on imported accelerators.

Alibaba has overhauled its Qoder coding agent into what it calls a universal workbench, aiming to move programming beyond professional developers and let non-technical users turn ideas into working software through conversation.

The new Qoder, unveiled on August 27, is built around coding as a core capability but is designed for everyone — from full-time developers to product, design, operations, data and independent creator roles. For non-programmers, code generation and execution run quietly in the background; users describe an idea in natural language and the agent breaks it down, refines the requirements and iterates through a chat-based workflow until a usable result is produced.

Alibaba frames the launch as a shift in the coding-agent market. Competition, it argues, has moved beyond pure code generation to agents that can autonomously complete broader work tasks. Qoder adds a desktop pet and real-time voice control so users can give instructions by speaking, with the assistant keeping track of running tasks. Two modes target different audiences: a programming mode for developers who want to select models and control the pipeline, and a general mode in which an auto-scheduling engine balances output quality, speed and cost without manual tuning.

Under the hood, Qoder builds on a year of development of its "Agent Harness" foundation, which the company says handles long-horizon tasks, decomposes complex work, executes code reliably and verifies outputs on its own. It also connects to a broad external ecosystem, including more than 40 connectors, 70 plugins and a 20,000-skill library spanning code repositories, project management tools, cloud services and internal business systems.

Alibaba is leaning on Qoder's fast-growing user base as evidence of traction. Since its first launch, Qoder reached 100,000 users in five days, passed one million in six weeks and generated $100 million in revenue within eight months. It now counts more than six million total users and over 100,000 enterprise customers.

The company also used the launch to offer limited-time incentives, including 500 Qwen model credits for individual users starting on five consecutive days through September 1, and a 50 percent discount for enterprises using the Qwen3.8-series models that power the new Qoder.

The broader bet is that coding agents become a mainstream productivity layer, not a developer-only tool. Alibaba's repositioning of Qoder is an attempt to claim that ground as the market shifts from code generation to autonomous task completion.

HarmonyOS, Huawei's self-developed operating system, has been formally recognized as the world's third-largest mobile operating system, and the company says its ecosystem is closing in on 100 million users by the end of this year.

At the HarmonyOS Ecosystem Conference 2026 on August 28, Huawei's rotating chairman Xu Zhijun revealed the latest milestones. Native HarmonyOS apps have surpassed 100,000, while more than 400,000 applications are available to users on HarmonyOS handsets. As of August 20, cumulative shipments of devices running HarmonyOS 6 had exceeded 80 million. Based on the current pace of weekly additions, Xu said the ecosystem should cross the 100-million-user mark in the fourth quarter.

The milestone matters because it signals the end of the "cultivation period" for the ecosystem. Once developers see enough users to justify investment, the flywheel of better apps drawing more users is expected to accelerate.

Recognition of its standing came two days earlier. At an August 26 State Council Information Office press conference, the Ministry of Industry and Information Technology said HarmonyOS had formally become the world's third-largest mobile operating system, a validation the company has been seeking since it decoupled from the Android-based line and pushed back-home devices onto its own platform.

Beyond mobile, Huawei is pushing the operating system into vertical industries. The conference announced the "Hongtu Plan," led by the Global Intelligent IoT Consortium (GIIC) and backed by Huawei, to drive chips, modules, operating systems and device makers toward broader commercial adoption across sectors. Devices built on OpenHarmony, the open-source variant, have now surpassed 1.35 billion cumulative units, with specialized versions deployed in markets such as automobiles, home appliances, metallurgy and petrochemicals.

Xu described the next phase as a test of global ambition. He said Huawei wants HarmonyOS to become a "second choice" for the world, and that the ecosystem must use a head start in AI to lead on intelligence before pushing overseas with open-source community partners.

The achievement comes as China's smartphone world consolidates around fewer operating systems, with HarmonyOS gaining ground on iOS and Android in the domestic market.

China has published a national standard that unifies the mechanical interface between flat-panel satellites and their launch vehicles, a step aimed at putting its low-Earth-orbit (LEO) constellations on an industrial footing rather than relying on bespoke designs for every batch.

The standard, GB/T 47811-2026 on the requirements for flat-stacked satellite rocket-satellite mechanical interfaces, was released on July 2 and takes effect on November 1, according to the Beijing Institute of Astronautical Systems Engineering. Drafted alongside China Satellite Network Group, the China Aerospace Standardization Institute, CAS Space and MinoSpace, it sets common rules for how multiple satellites are stacked inside a rocket fairing, secured before liftoff and separated in orbit one by one.

The point is interoperability on a scale that has historically been missing. When satellites used their own load paths, mounting dimensions and stacking heights, every new mission forced a rocket to be re-engineered. With a shared interface, the same batch of satellites can in principle switch launch vehicles, and rocket makers no longer have to design a fresh installation scheme for each task.

The timing tracks a period of rapid ramp-up. On August 6, 2024, a Long March 6A lifted the first 18 Qianfan satellites, marking China's first "18-in-one" flat-satellite stacking launch. In July this year, a Long March 6A raised the Qianfan fleet to 218 satellites, and the next day a Long March 8A flew with 20 — the first jump from 18 to 20 stacked satellites in a single mission. As of August 27, the Qianfan constellation had deployed 238 satellites in total.

The new interface rules absorb lessons from those launches, as well as from the separate China Satellite Network Group program. The logic is straightforward: run enough launches to gather real engineering experience, then turn that experience into a repeatable rule so production and delivery can become continuous.

CAS Space, one of the standard's drafters, will be an early test of the route. Its Lijian-2 rocket completed a first flight on March 30, and a second Lijian-2 mission is planned for October to carry 18 satellites into orbit. That mission, contracted to the Shanghai Spacecom Satellite Technology-led Qianfan program, has not been confirmed to use the new standard, since it launches before the rules take effect. But it sits on the same path toward industrial-scale constellation assembly that the standard is meant to accelerate.

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.

The 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.

Hunyuan 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).

This 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.

Positioned 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.

In 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.

In 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.

In 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.

In 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.

Notably, 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.

Hy4 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.

Hy4 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.

Since 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.

Since 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.

With 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.

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.

Usage 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.

A 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.

The 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.

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.

Today 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.

What 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.

Hong 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.

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.

Baidu 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.

The 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.

Behind 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.

The 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.

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.

Three 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.

XPeng 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.

The 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.

For 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.

The line that once defined a premium large language model has shifted almost overnight in China. On a single night, Zhipu AI released and open-sourced GLM-5.3-Flash — confirming it was the anonymous 'Ox Alpha' that had topped OpenRouter since August 20 — while Alibaba rolled out Qwen3.8-Flash, sharing weights on Hugging Face and its ModelScope community. Both are 'Flash' editions, and both are aggressively priced, upending the convention that more capable means more expensive.

GLM-5.3-Flash keeps a one-million-token context window and is the first natively multimodal model in the GLM-5 family, handling images, video, files and even computer use. During its limited-time discount, input costs 0.4 yuan per million tokens and output 1.4 yuan, roughly one-twentieth the price of GLM-5.3. Qwen3.8-Flash, positioned as an early preview of the Qwen4 architecture, also retains a million-token window; at 1 yuan per million input and 3 yuan output it costs about one-twelfth of Qwen3.8-Max, while cutting token consumption by 75% in Qwen office scenarios.

The comparative rapid-fire puts pressure squarely on DeepSeek, long China's price-performance benchmark. DeepSeek V4 Flash currently lists at 1.5 yuan input and 4.5 yuan output per million tokens, with peak and off-peak multipliers. In benchmark testing, GLM-5.3-Flash performed close to Opus-class models — a level that Flash labels us to expect — and its anonymous run ended DeepSeek's 56-day reign at the top of OpenCode. Zhipu said the traffic was served by more than 100,000 domestic AI chips, which it claims reached hardware efficiency comparable to mainstream NVIDIA GPUs.

The deeper signal is strategic. 'Flash' is shedding its meaning as a stripped-down bargain. Builders now frame these models as the real flagship: ample context, multimodal input, competitive reasoning and a price that makes them the default for everyday production loads across coding agents and office suites.

The result is a market where price anchoring is fragmenting fast. Model vendors are no longer competing only on ceiling capability, but on serving practical workloads cheaply enough that most developers never reach for a 'Pro' tier. In China's LLM market, cheap and generous has quietly become the new flagship bar.

After holding through tough competition, Xiaomi finally stabilized. On August 18 it reported 2026 second-quarter results: revenue reached 108.92 billion yuan, again above 100 billion, with adjusted net profit of 6.22 billion and attributable net profit of 9.46 billion, doubling quarter over quarter. Phones and AIoT remain the mainstay at 84 billion, with smartphone revenue 42.1 billion. Autos are the other big line: Xiaomi delivered over 100,000 vehicles, up nearly 30% year over year, for 23.9 billion, and with AI and innovative businesses the segment reached 24.9 billion. By Omdia counts, Xiaomi shipments ranked top three in 53 countries; the AIoT ecosystem has 1.16 billion connected devices, up 17.4%, with 24.6 million users holding five or more. Compared with Q1, Xiaomi walked out of its darkest moment: phones and autos have stabilized.

The real opportunities sit in the new wave of technology. Xiaomi kept high-intensity R&D: Q2 spending reached 9.2 billion, up 18.9%, with at least 200 billion planned over five years. The spend goes mainly to AI and chips. Lei Jun called AI the next decade's largest opportunity, with 2026 AI R&D and capital exceeding 16 billion and 60 billion over three years. Returns are emerging: in Q2 Xiaomi released the MiMo-V2.5 model, topping both OpenRouter's monthly and weekly call charts over DeepSeek-V4 Flash. CFO Lin Shiwei said the result rests on advantages in model capability, inference efficiency, and cost control, and a new MiMo model is training for imminent release. HyperOS 4 gives MiMo scenarios across phones, IoT, and cars, with Super Xiaoai 2.0 enabling cross-app, cross-device task orchestration. MiMo will launch its first desktop application, a market with players including Doubao, Kimi, and QwenWork. The autos-and-AI segment's other business income was 1 billion, up 56.5%.

Chips are another focus. Lei Jun declared a 50 billion, 10-year push to build Xuanjie chips; president Lu Weibing said the O1 has shipped over one million units across three terminals, with a new-generation chip coming. In robotics, Xiaomi released two embodied foundation models ranking first on WorldArena and RoboCasa, and its new humanoid entered factory work at 98% success on a self-tapping nut workstation. From foundation models to applications, chips to robots, Xiaomi spans the hot tracks.

The new assets still need investment, while phones and AIoT seek breakthroughs. With chip-price increases and weak demand, global smartphone shipments fell about 6%, per Omdia. Xiaomi's phone strategy is premiumization and globalization: average selling price rose to 1,351 yuan, up 25.9% to a record, with domestic models at 3,000 yuan or above at 32.1% of sales, a record, and overseas premium share also at a record. The worst is passing and the inflection point arrives in the second half.

Alibaba's Qwen App released a major update on August 7, bundling office assistant, scheduled task and flagship Qwen3.8-MAX model access into a single version of the consumer AI app. For the first time, the release carries an explicit pricing signal: core features stay free, but the new office assistant will run on a quota that requires payment once exhausted.

The pattern is the same one that ByteDance's Doubao set in June, when it launched a three-tier subscription ranging from 68 to 500 yuan per month. Tencent followed with its own personal tiers between 99 and 999 yuan, and Alibaba itself quietly shipped a 98-yuan-per-month standard subscription for its corporate QwenWork product on August 3. Office workflows have emerged as the segment where Chinese AI vendors believe users will actually pay.

Qwen App sits inside a more awkward position than its peers. It was built as Alibaba's AI-to-consumer flagship, integrated with Taobao, Alipay, Fliggy and Amap, and fronted a 30-billion-yuan New Year marketing campaign that pushed its daily active users from 7 million to a peak of 73.5 million. Once the campaign ended, DAU fell by nearly half. Per-user session time also collapsed, from 6.3 minutes before the campaign to a low of about 3 minutes after, a 51% drop. Customer acquisition cost ran roughly 144 yuan per DAU, well above ByteDance's Doubao and Tencent's Yuanbao.

Internally, Qwen App also collides with the QwenWork product that Alibaba launched publicly on August 3 under the new DingTalk CEO Chen Yusen. The office product consolidated three previously competing agent products, QoderWork, Wukong and MuleRun, into a single entry point. With DingTalk handling enterprise sales and QwenWork absorbing the productivity roadmap, Qwen App is left to argue for a distinct consumer identity it has not yet built.

The paid-office bet is therefore less about monetization than about narrative. Whether Qwen App can clear enough quota usage to qualify as a real office product will determine whether it ends up as the consumer front-end to Alibaba's AI stack or simply as another chatbot window in a crowded market.

China's AI office track reached 30 million total monthly active users on desktop PC clients in July 2026, according to the first AICPB AI Product Ranking desktop list. Tencent's WorkBuddy led the overall chart with 11.15 million MAU and growth of 304.4%; Baidu Dazi had 6.74 million MAU with growth of 1,063.79%, first on the AI office agent growth chart; and ByteDance's TraeWork had 2.66 million MAU with growth of 51.08%. Newer entrants include Alibaba's QwenWork at 607,900, Nano Work at 227,300, and Kingsoft Lingxi at 258,200 at the end of July. The AI office war has officially begun, with BAT and ByteDance moving quickly and updating products almost daily.

Baidu released Dazi in March, integrating Baidu Baike, Baidu Maps, MiaoDa, and cloud drive capabilities into a professional office agent platform. In July it reached 6.74 million MAU at 1,063.79% growth. Alibaba's QwenWork, consolidated at end-July, absorbed Wukong incubated by DingTalk, MuleRun started internally by Alibaba Cloud, and QoderWork, with July MAU of 607,900 for QwenWork, 588,500 for Wukong, and 1.89 million for QoderWork. Tencent's WorkBuddy grew from a small team to substantial resource support, connecting Tencent Docs, QQ Mail, Tencent Drive, and Tencent Meeting lines, reaching 11.15 million MAU, first overall, at 304.40% growth.

ByteDance merged the Feishu product team into Doubao, launched work tasks, and connected Feishu Sheets and other office capabilities. Because Doubao's entertainment positioning is strong, ByteDance is also pushing TraeWork and Coze as AI office products, with TraeWork at 2.66 million MAU at 51.08% growth and Coze at 2.17 million at 62.23%. Other players include NetEase Youdao's LobsterAI at 511,100, SenseTime's Xiaohuanxiong at 379,700, Cheetah's easyclaw at 274,900, Kingsoft's Lingxi Pro at 258,200, 360's Nano Work at 227,300, Zhipu's AutoClaw at 143,500, iFLYTEK's Loomy and AstronClaw at 7,400 and 1,400, and Kuaishou's krowork at 3,000.

While the ranking does not freely publish DAU data, it offers an estimation method: for good AI office agent products, DAU is roughly equal to 0.2 times MAU. With AI office MAU now at only 30 million, compared to the user scale of traditional office software, at least 20x more headroom remains. This huge growth space suggests competition in AI office will intensify; facing the entrance dividend of 600 million monthly active AI users, no player can afford to ignore the market, and once lost, the entire AI era could slip away.

For users, whoever ultimately wins, the practical benefit of competition is rapid product upgrades and token discounts. The ranking notes its data is based on publicly available information with statistical limitations, and that non-commercial citation should credit AICPB.

ByteDance founder Zhang Yiming made a rare appearance at the company's Beijing headquarters two weeks ago and sent the Seed AI research team a message that amounted to a public line in the sand. Distillation, Zhang told the team, is off the table. The shortcut of using a larger competing model to train a smaller one can lift benchmark scores quickly, but it caps long-term technical breakthrough. Build real capability, Zhang said, or stop pretending you are building AI.

The remark matters because distillation has been the silent backbone of the Chinese open-source model wave for the past eighteen months. Renting API calls from a frontier foreign model, generating labeled corpora at scale, and post-training a smaller model on their outputs has been the cheapest path to a top-ten OpenRouter ranking. Zhang Yiming has now noticed, and he has put it on the record inside one of the largest AI research teams in China.

The mid-year All-Hands on August 6 made the structure change explicit. CEO Liang Rubo announced that Doubao and Douyin are now the company's two 'thick main lines'. Doubao is no longer a model wrapper; it is a flagship business line on par with Douyin. Zhao Qi, the Doubao product lead, has been elevated to the main table for the first time. Feishu product and Doubao product are merged into a single Doubao team, with former Feishu head Xie Xin now reporting to Zhao. Volcano Engine and Feishu GTM are merged into a Creativity Service Platform under Tan Dai. It is ByteDance's largest B-side restructure since the 2021 BU overhaul.

The strategic logic has shifted. The All-Hands positioned AI as an AGI bet, not a feature race. Liang called out the 'China-style innovation' framing directly, arguing that the next generation of productivity infrastructure will be defined by who owns the agent entry point into the enterprise. Alibaba consolidated QoderWork, Wukong and MuleRun into QwenWork in recent months. Tencent has elevated its SkillHub to a top-level strategy. ByteDance's answer is to merge Feishu into Doubao and let the agent be the entry point.

The financial logic is just as pointed. LatePost reported that Doubao's daily revenue was under one million yuan in the first half, while daily compute cost ran into tens of millions. The June launch of a paid Pro tier at 68, 200 and 500 yuan per month was the first commercial experiment. Zhang Yiming's distillation ban is the long-game version of the same math. Cheaper models that win on rented capability cannot carry the B-side compute bill. AI for ByteDance now means building it, not borrowing it.

On August 25, ByteDance launched Doubao Work, an AI office product spun out of the work-task mode inside its Doubao assistant. Now a standalone app with its own desktop client — while still embedded in Doubao PC — it can draft documents, build spreadsheets and presentations, generate images, video and web pages, and even operate a browser or cloud computer to research and fill forms.

It enters a crowded field. Tencent's WorkBuddy launched in March, Alibaba consolidated QoderWork, Wukong and MuleRun into QwenWork in early August, and Kingsoft keeps pushing WPS Lingxi. Overseas, Anthropic shipped Claude Cowork early this year and OpenAI folded Codex into ChatGPT Work when it merged the two desktop products in July. ByteDance, Alibaba and Tencent have all been reorganizing teams around AI office in recent weeks.

Doubao Work's clearest differentiation is its tie to Feishu, ByteDance's workplace collaboration app. Signing in with a Feishu account lets the assistant pull chat records, documents, meeting notes, calendars and permissions within scope, complete tasks, and write results back. For ByteDance, this gives the product a ready-made work memory covering the context generic agents lack — what was discussed last week, which version the boss changed, where a project stands.

The launch also mirrors a deliberate split between chat and work. Just as ChatGPT separates the two, Doubao Work gets a standalone client rather than living only inside an input box, giving longer-running office tasks a dedicated workspace of skills, connectors and cloud computers.

The real battleground, analysts note, is which office ecosystem a customer already lives in. Companies deep in Tencent's tools find WorkBuddy most convenient, DingTalk users gravitate to QwenWork, Feishu-centric teams suit Doubao Work, and WPS loyalists stay in WPS Lingxi. Models swap quickly; the hundreds of thousands of documents, years of meeting records and permission systems do not. For now, Doubao Work bets that knowing what you are working on is worth more than having the smartest model.

Late on August 26, Alibaba’s Qwen team released Qwen3.8-Flash, an open multimodal mixture-of-experts model that it describes as a technical preview of the next-generation Qwen4 architecture. The name still carries a 3.x tag, but the design is a new generation, released early so the community can validate the architecture ahead of the full Qwen4 family.

The headline is cost. The main model has 125 billion parameters with an additional 51 billion-parameter N-gram embedding module, but only about 6 billion parameters are activated per token. Where a dense model touches every parameter for each token, the MoE design calls only the relevant experts, keeping compute bills low. Native context spans 262K tokens, extensible to 1 million via YaRN.

Alibaba said training cost is about one-ninth that of Qwen3.7-Plus, yet coding and office tasks are stronger. Scores include 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro and 84.5 on AndroidWorld, clearing the bar set by the value benchmark DeepSeek V4 Flash. API pricing follows the low cost: 1 yuan per million input tokens and 3 yuan per million output tokens — roughly a third of V4 Flash, with no peak and off-peak distinction.

Architecturally, attention uses a GDN-plus-QSA hybrid — GDN compresses historical information while QSA filters important context at fine granularity through a lightweight indexer. On 1-million-token sequences, the QSA attention kernel reaches up to 7.6x and 4.9x speedups in prefill and decode. The N-gram embedding can be offloaded to host memory and prefetched asynchronously, and training uses the Muon optimizer with a refit scaling law.

Weights opened at 11pm Beijing time on Hugging Face and ModelScope, alongside an FP8 quantized version that lowers the barrier to local deployment. Alibaba released it as Qwen3.8-Flash-Next, explicitly framing it as a Qwen4 blueprint meant to gather community feedback before the full family arrives, and it ships in the Qwen AI platform with a default 1-million-token context and built-in tools, debuting in a new office product.

For developers, the early release is a chance to adapt to the Qwen4 architecture ahead of the full family and gain first-mover advantage, while the aggressive pricing — output at roughly one-thirtieth of many frontier overseas models — opens up use cases that were previously uneconomical, from long-document processing to agent workloads and batch data handling.

Chinese tech giant Huawei and U.S. PC maker HP Inc. jointly announced on August 26 that they have signed a multi-year global patent cross-licensing agreement. Under the deal, HP gains the right to use Huawei’s Wi-Fi-related patents across computers and other electronic peripherals it sells worldwide.

Huawei chief IP officer Fan Zhiyong said in a statement the company was “pleased” to reach the agreement, calling it strong proof of Huawei’s continued independent innovation in frontier information and communications technology. Steven Geiszler, who represented Huawei in the talks, noted the cross-licensing deal includes a valuable reciprocal patent grant from HP.

The two companies were not starting from scratch. In August 2025, Huawei sued HP in the Munich division of the European Unified Patent Court over an unlicensed Wi-Fi 6 patent. In November 2025, HP licensed into the Sisvel Wi-Fi 6 patent pool, gaining one-stop access to about 2,000 standards-essential patents, after which Huawei dropped its lawsuit.

The new agreement extends Wi-Fi 6 rights and broadens coverage to Huawei’s wider Wi-Fi portfolio, including the latest Wi-Fi 7 standard. An HP spokesperson stressed the deal is a standards-essential patent license for Wi-Fi technology and does not represent a broader strategic or commercial relationship with Huawei.

The pact underscores Huawei’s depth in wireless communications even as the U.S. government continues pushing allies to restrict its business. Huawei said more than 1.6 billion consumer electronic devices (excluding phones) used its Wi-Fi patents by the end of 2025, from licensors including HP, Microsoft, Sony, Asus, Acer and Amazon. Huawei charges a $0.50 per-unit royalty on consumer Wi-Fi 6 and Wi-Fi 7 products and reported $630 million in patent and royalty income for 2024.

The agreement is the latest example of Huawei’s patent-operations strategy expanding amid a complex international environment. Huawei said it has signed more than 260 patent licensing agreements with major ICT companies across the U.S., Europe and Japan and South Korea, adding 34 new ones in 2025. Beyond building revenue from its portfolio, the deals signal that Huawei’s wireless technology remains widely adopted outside China even where its equipment business faces restrictions.

On August 26, Zhipu AI confirmed that the anonymous model Ox Alpha — known as “Niu Lai” in Chinese developer circles — is its newly released GLM-5.3-Flash. Before its official debut, the model was offered for free in anonymous testing on OpenRouter and OpenCode, drawing more than 50 trillion tokens of traffic in five days and setting growth records on both platforms. Usage has since been more than double that of DeepSeek.

What drew market attention was a detail Zhipu disclosed afterward: all of that traffic was served by domestic chips. In its technical documentation, Zhipu said its inference service ran on a cluster of more than 100,000 domestic chips and that “hardware efficiency and per-token cost have reached a level comparable to mainstream Nvidia GPUs.”

Semiconductor research firm SemiAnalysis commented on X that with every request carried by domestic silicon at Nvidia-comparable efficiency, “the CUDA moat is once again being tested,” following OpenAI’s announcement of its in-house Jalapeño inference chip. LatePost reported the chips may come from Huawei, Moore Threads and Hygon, though Zhipu would not confirm the suppliers or models.

To scale toward a 1-million-token context, Zhipu built a dedicated inference engine on SGLang using W8A8 quantization, mixed INT8/FP8/BF16 cache quantization and node-level tensor parallelism, plus a production Encode-Prefill-Decode separated architecture that splits multimodal encoding, prompt prefill and token-by-token decoding into independently schedulable pools. The company says end-to-end service performance improved 3x over its initial baseline on the same hardware.

GLM-5.3-Flash has 320 billion total parameters with 18 billion activated, roughly matching DeepSeek V4 Flash. It scores 57 on the Artificial Analysis Intelligence Index, on par with Claude Opus 4.8 and above DeepSeek’s V4 Pro. Pricing is about one-tenth of GLM-5.3 and one-fortieth of Claude Opus 4.8, positioning it as a low-cost option amid a wider round of Chinese model price increases.

2026 is being called China's mass-production year for humanoid robots. At the World Robot Conference, the numbers looked stirring. Counterpoint Research says global humanoid shipments surpassed 22,000 in the first half of the year, up nearly 300% year over year, while the Ministry of Industry and Information Technology expects China's full-year output to exceed 100,000 units.

The top Chinese makers are scaling hard: AgiBot shipped roughly 9,700 units in the first half, Unitree about 7,000 by August 21, and Galbot more than 1,100 over the first half. UBTECH announced all-channel orders for its full-size superhumanoid U1 series exceeding 13,361 units as of June 30, aiming to deliver this year — though founder Zhou Jian warned the manufacturing difficulty 'is rare in human history,' with 2,000 to 3,000 head components alone.

Beneath the headline volume sits a stubborn reality. While entertainment and performance plus education and data collection still made up over 60% of global humanoid shipments in the first half, smart manufacturing accounted for just 13% and warehousing and logistics for 5% — meaning fewer than two in ten units shipped are actually going to work in factories and warehouses. Unitree's prospectus shows more than 70% of its humanoid-related revenue in the first nine months of 2025 came from R&D and education, against about 9% from real industry applications.

Still, a few operators have pushed past demos. One embodied-AI firm claims to be the first to reach product-market fit in logistics, running parcel-sorting lines with SF Express and China Post across more than ten logistics centers in five provinces — drawn by night-shift environments that are un-air-conditioned, noisy and physically grueling, yet offer clear metrics for payback.

Executives frame the challenge as three gates: making a robot work once, making it repeat the task reliably, and turning a one-off project into a reusable product. Bank of Communications International argues the industry is shifting from showing capability to proving value, with commercial evaluation moving toward efficiency, stability and ROI. The real test of the mass-production year, analysts say, is not how many units leave the factory, but how many are asked to come back to work.

Hours before the World Robot Conference closed, the Beijing Humanoid Robotics Innovation Center swept the medal table at the second World Humanoid Robot Games, taking 15 gold, 12 silver and 18 bronze medals by 10 p.m. on August 26. Its Tiangong series set 34 human records during the games and took gold in the competition and scenario headliner events.

The center had a busy week. At the conference it unveiled Pelican-Unify, a unified embodied-intelligence model, and Tiangong Omni, a lightweight humanoid standing 1.35 meters and weighing 39 kilograms, aimed at lower-cost mass production. It also announced a deep strategic partnership with Mercedes-Benz, which put Tiangong 3.0 to work as a 'silicon-based recommender,' greeting visitors, explaining models and fielding questions in a dense crowd.

CEO Xiong Youjun argues the industry has reached the 'eve of embodied AI's ChatGPT moment.' Competition, he says, has moved from chasing headline parameters such as motor torque and degrees of freedom toward systematic capability — supply chain, engineering and customer experience — and toward the return-on-investment question that marks a maturing sector. He acknowledges humanoids still trail industrial robots and humans on ROI, but insists the gap is closing fast.

Xiong sees the industry's biggest problem as waste: duplicated data platforms, high verification costs and incompatible technical stacks. His answer is a common infrastructure layer. Tiangong serves as a general-purpose body platform, while its open-source Huisi Kaiwu provides an operating-system-level base that has logged more than 16 million downloads, ten-plus open models and over 200 secondary-development partners, including institutions in the United States and Europe.

The pathway is deliberately unglamorous. Beijing Humanoid is focusing on what Xiong calls 3D scenarios — dirty, dangerous and dull work such as live-line operations, laboratory testing, energy inspection and logistics sorting. It helped the Jiangsu Institute of Metrology automate infrared-thermometer verification six times faster. Xiong is also pushing standards, chairing the safety-standard group and advancing 49 standards. 'Before an industry truly runs,' he says, 'someone has to pave the road first.'