ByteDance's AI office app Doubao Work has added a creation canvas for complex design work and a lighter model, Doubao 2.1 Lite, for everyday office tasks, Leiphone reported on October 9.

The canvas puts source materials, design options and finished output on one infinite page, so users can carry out each step of a project in one place and view, compare and organize results as they go. After an image is generated, users can keep editing its text, colors and layout, adjusting from overall style down to local details. The canvas is wired to Seedream 5.0 Flash, a new ByteDance image model. In complex creative workflows, the app's agents can start from understanding the request and planning an approach, then produce images, videos and full sets of designs.

To use it, users open Doubao Work, or the work-task mode in Doubao's web and desktop versions, tap Skills and choose Create Canvas, or open the Creative Design skill. Anything generated from a prompt flows onto the same canvas. Doubao already sells a creation credit pack with higher usage limits for people who do a lot of design work.

Doubao 2.1 Lite is aimed at high-frequency office work such as everyday questions, document writing, spreadsheet handling and slide decks, and is tuned for faster responses and lower credit consumption. ByteDance says it is relatively strong at understanding images, video and audio, with better results in multimodal education and office scenarios. Users can select it after updating the client to the latest version.

The release keeps up a weekly update cadence that Doubao Work has followed since its launch on August 25. Earlier updates added parallel multi-agent work and computer control on Mac, local Office file editing, browser screen recording and playback, a goal mode and a plan mode, as well as an upgrade to the Doubao 2.1 Pro model.

The canvas moves Doubao Work further into visual design, an area where office agents have so far focused mostly on documents, spreadsheets and slides. The announcement did not include benchmark results for Doubao 2.1 Lite or Seedream 5.0 Flash.

SEE ALSO: Doubao Work Adds Parallel Multi-Agents and Mac Local GUI Control · ByteDance Launches Doubao Work to Take on QwenWork and WorkBuddy

KingKong Technology, a Beijing company best known for robot encoders and servos, has open-sourced the software stack behind Jumper, a small crab-style robot it designed itself. The release covers the robot model, simulation scenes, reinforcement learning training, motion control and deployment tooling, letting developers retrain the walking, dancing and jumping moves the company showed in its demos.

Jumper has 22 degrees of freedom on six legs, measures about 400 by 400 by 200 mm and weighs about 1.8 kg. The company lists a maximum jump height of 400 mm or more, a top speed of at least 0.5 meters per second, a 1 kg grasping load and about two hours of battery life. Its hardware page lists a Rockchip RK3576 processor with a 6 TOPS NPU. Each of the 22 joints uses one of the company's tactile servos, which carry two encoders to read both the motor side and the joint side after the reduction gearing.

The open-source repository ships Jumper's MuJoCo model and training environments, built on MuJoCo, MuJoCo Warp, mjlab and rsl_rl, so the same tasks, rewards and PPO code can run on GPU or CPU physics backends. Example skills include walking, posture changes, gestures, dancing, jumping and grasping, along with several hexapod gaits. Developers can change task configurations, reward functions and control parameters, then retrain.

Trained policies can be exported to ONNX, converted to RKNN for the onboard chip and packaged with Jumper's controller and state machine into an app-style motion bundle. The project also expects developers to work through an AI coding assistant, describing new appearances, motions or scenes in plain language.

The openness has limits. Mechanical CAD files, the production bill of materials, PCB designs, electrical schematics and full servo manufacturing data are not public, so developers cannot rebuild an identical robot from scratch. The documentation also states that RKNN inference on the real board has not yet been validated and that the 1 kHz controller loop does not yet drive the physical motor bus.

For KingKong Technology, which has mostly supplied components to other robot makers, Jumper is a first step into defining its own robot body, software and skills. The company is running an open motion challenge for developers and pitches home chores, pipe inspection, sorting and farming as possible uses, though which of these have real demand remains to be tested.

SEE ALSO: Micbot Says Robot-Dog Delivery Dispute With Zhejiang Buyer Settled After Delay and Dev-Rights Clash · Galbot and Tsinghua's LATENT Wins an IROS Award for a Humanoid Tennis Forehand

ByteDance's AI development brand TRAE has merged its two clients, TraeWork and TraeCode, into a single product. Announced on October 9, the new TRAE uses one entry point and lets users switch between an Agent mode and an IDE mode, a setup the company describes as a full-chain development platform covering everything from breaking down requirements to delivering finished work.

Until now the two apps served different users. TraeWork leaned toward office tasks, task management and content generation, while TraeCode focused on professional developers' coding needs. Under the merged product, Agent mode takes a task from an idea to a delivered result. Beginners can build a demo step by step, while experienced developers can run multiple agents in parallel across projects. IDE mode keeps the full intelligent development environment for deep editing and fine-grained debugging within a single project.

Features from both apps now sit in one workflow. Automated tasks hand recurring jobs such as progress summaries, routine checks, test-result roundups and report generation to agents that keep running them. A template library speeds up common starting points such as product proposals, page builds, data analysis and technical documentation. My Assistant brings non-coding work such as documents, reports and collaboration into the same task chain. Outputs including code changes, HTML, slides, Word documents, images, videos and reports are stored together under Project Artifacts so they can be reviewed, shown and iterated on.

The new TRAE runs on desktop, web and mobile. A user can start a task on a computer, check progress and add requirements from a phone, then return to the desktop for deeper development, with projects, tasks and artifacts staying in sync across devices.

The merger comes with a deadline. TRAE's documentation says TraeWork will stop service on October 23, and users who have not upgraded by then will lose access. Chat history, account data, local data, login status and commercial credits are migrated automatically; TraeWork users start the upgrade manually in the client, while TraeCode users receive it as an automatic update.

Early reaction is mixed. Some developers in TRAE's official community have complained that the upgraded app opens into the Agent workspace by default, that IDE mode must be selected by hand, and that there is not yet an option to change the default startup mode.

SEE ALSO: ByteDance's TRAE Work Adds Design Mode, Closing the Loop From Requirements to Code · ByteDance Launches Standalone Version of TRAE AI Coding Tool “SOLO”

openJiuwen, the open-source AI agent platform built by Huawei 2012 Laboratories, Huawei Cloud and other Huawei teams together with universities and companies, has released and open-sourced an enterprise-grade AgentOS. The project targets the gap between agent demos and production use: long task chains that several agents struggle to coordinate, experience that is lost between runs, rising compute spend, and the data permissions, isolation and recovery that enterprises need before they trust agents with real work.

The code is published as AgentOS for Enterprise on GitHub and AtomGit under the Apache 2.0 license. Its repository describes four components: a distributed agent runtime, the jiuwenswarm work and coding agents, the Conch agent sandbox and an agent access gateway. It supports single-machine and cluster deployment, with one command to build and deploy, and runs on openEuler or Ubuntu.

On coordination, AgentOS uses swarm collaboration and workflow orchestration to split tasks, assign roles and run steps in parallel, for example research, analysis and writing, with people able to step in at key decision points. On self-evolution, it uses execution traces, user feedback and memory to refine agents' skills and collaboration patterns, turning what worked into reusable capabilities.

The runtime also aims to use compute more efficiently through hardware-software co-design, resource scheduling and inference optimization. For safety, it offers multi-tenant isolation, security sandboxes, permission management and guardrails that limit data access and tool calls, plus fault recovery so long tasks can keep running.

Model calls, tool calls and memory management are packaged as plugins that load on demand. For developers, five kinds of assets, namely skills, connectors, plugins, experts and expert teams, can be combined and published, obtained or shared through Agentic Hub, so partners can package industry methods and system integrations as reusable assets.

The software is also the base for hardware delivery. At Huawei Connect 2026, Huawei launched an open-source agent acceleration platform for all-in-one appliances built on Jiuwen AgentOS. It bundles the unified agent gateway, the distributed runtime and the WorkSwarm office and coding workbench, with a built-in SystemAgent for one-click setup and management. According to the launch information, enterprises can complete an end-to-end private deployment in hours, then connect existing business systems, local knowledge bases and industry models through open plugins and interfaces.

SEE ALSO: Huawei's openJiuwen Gives WorkSwarm Full-Duplex Voice and Video, Ready to Deploy on Ascend NPUs · Huawei Cloud Rolls Out AgentArts and Agentic Cloud Stack for Enterprise AI at Connect 2026

Huawei's LogicFolding chips will not stay confined to a handful of flagship devices. Richard Yu, Huawei executive director and chairman of its Consumer Business Group, told an international media roundtable that "there will definitely be more Huawei phones" using logic-folding chips built under the company's Tau (τ) Scaling Law, according to IT Home and TechWeb. The technology entered Huawei's phones with the Mate XT 2 tri-fold, and Yu said it will gradually move into more of the company's smartphone lines.

Yu gave no model names or timetable for the wider rollout. He tied the pace partly to manufacturing, saying he hopes China's semiconductor industry can further expand capacity so Huawei can build more advanced products for consumers at home and overseas.

At the same roundtable, Yu laid out which parts of the Kirin 9050 Pro, the chip in the top Mate 90 models, Huawei designed itself. The list covers the Balong modem, the image signal processor (ISP), the mobile security processor (MSP), the Lingxi CPU, the Maleoon GPU and the NPU built on Huawei's Da Vinci architecture. Yu said the phone's overall performance is 31% higher than the previous generation, and that in-house design delivered better energy efficiency and system performance despite limited access to advanced manufacturing processes.

The rollout comment comes with a forward roadmap. Under the Tau Scaling Law path cited in the reports, Huawei's 2026 chips reach a performance-core frequency of 3.1 GHz, a 12.7% step up from the 2.75 GHz of the Kirin 9030 Pro. Frequency and transistor density are then set to rise steadily, reaching more than 400 million transistors per square millimeter and a 5.0 GHz clock by 2031.

LogicFolding stacks a chip's logic layers vertically and uses dense interlayer connections to shorten critical paths and cut delay, rather than relying only on smaller transistors. Huawei has not said which product lines will adopt it next, or whether mid-range phones are included.

SEE ALSO: Huawei Ships Kirin 9050 Pro With LogicFolding Architecture Under Tao's Law · Richard Yu Says Huawei Has Mass-Produced 381 Tau Chips, From Phones to Smart Cars

The Centre for Artificial Intelligence and Robotics (CAIR), Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, has unveiled CARES 4.0, a multimodal clinical AI agent system. It was launched at the 4th CREATE Symposium on clinical robotics and embodied AI, organized by CAIR with Hong Kong Science and Technology Parks Corporation at Hong Kong Science Park. CAIR says the agent framework and general toolkit have been released, medical toolkits are rolling out, and its ultrasound and surgical-video models will join the platform free of charge.

CAIR director Liu Hongbin told Leiphone the change is in kind, not quantity. Versions 1.0 through 3.0 were a model that answered questions about images or videos, a tool doctors had to operate themselves. Version 4.0 is an agent that carries out tasks. In his example, a department head who wants last month's surgical volume, consumables and patient mix in a report once needed an assistant working with CARES 3.0; the agent can now assemble it, checking back on what kind of analysis is wanted.

The system is built on the Harness agent framework and CAIR's own foundation models for CT, MRI, ultrasound, endoscopy and EEG. CAIR says it has been validated at several top-tier hospitals in three areas: assisting the diagnostic workflow from complaint through imaging and pathology, answering natural-language questions about department operations such as equipment load and case mix, and running statistical analyses and simple machine-learning models on specialty datasets.

Liu stressed that CARES acts as an assistant and gives suggestions. He acknowledged that AI will sometimes hallucinate and said the first priority is to minimize that, especially for decision-type advice. The system is designed to follow clinicians' reasoning, drawing on guidelines and recent literature organized as structured data, and each suggestion comes with its reasoning path and sources. "The doctor is the final decision-maker," he said.

One function already in clinical use is spatial risk warning during procedures, alerting a physician when an instrument nears major vessels or nerves. In bronchoscopy at the First Affiliated Hospital of Sun Yat-sen University, it can flag that an apparent bump on an airway wall is actually part of a blood vessel, where a biopsy could damage an artery.

Liu said CAIR's systems, mostly earlier tools, are in use at more than 30 hospitals, including laryngoscopy workstations at more than 10, bronchoscopy at more than four and ultrasound at more than 10. Doctors' most common requests, he said, are less effort and clearer displays so AI output is not misread.

SEE ALSO: Alibaba DAMO Academy Open-Sources Expert-Level Abdominal CT Model DAMO RADAR in Science · United Imaging Intelligence Meta Hospital Strategy: Future Hospitals Without Walls as AI Agent Orchestrates Care Beyond Physical Boundaries

Huawei has moved DevEco Studio for HarmonyOS PCs from invitation-only testing to an open public beta, letting developers build and debug HarmonyOS apps on Huawei's own PCs. The company announced on October 8 that two companion tools, the AI development tool DevEco Code and the agent toolkit DevEco CLI, are entering their first public betas on HarmonyOS PCs at the same time.

The beta runs on several MateBook models and the MatePad Edge, and requires HarmonyOS 7.0.0.107 or later, at least 16 GB of memory and at least 100 GB of storage.

The IDE is built on Huawei's in-house BitFun platform, which the company credits for low resource use and fast startup. It covers code editing, completion, semantic analysis, builds, debugging and performance tuning, with plugin support and multi-module project management. Developers can launch and debug apps in seconds without installing an emulator, switch the debug window between phone, tablet and foldable layouts, or open several windows so one code change can be checked across form factors. A two-way preview links the component tree to the matching source code.

DevEco Code is the AI layer, and its capabilities are also embedded in DevEco Studio. It ships with Zhipu AI's GLM-5.3 and GLM-5.1 models, uses a harness that Huawei says raises compile pass rates and task completion for HarmonyOS apps, and lets users connect third-party models. It draws on a HarmonyOS knowledge base, Skills and MCP tools, and offers three modes. In Plan and Build, the developer reviews a plan and approves its execution. Goal generates code from a spec document, then verifies and fixes it until the target is met. Huawei says the tool can compile, run and functionally check code on its own.

DevEco CLI opens these capabilities to other agent tools. A single init command deploys DevEco CLI Skills into multiple agent tools on the same machine. The toolkit includes command-line tools for project setup, syntax checks, builds, managing physical devices, pushing packages and capturing debug logs; an official knowledge base of guides, API references, best practices and change notices that Huawei says is manually calibrated; and more than 70 Skills distilled from HarmonyOS development practice and bug fixes.

Together, the three tools give HarmonyOS PCs a native development toolchain, which Huawei describes as developing HarmonyOS on HarmonyOS.

SEE ALSO: Huawei Cloud CodeArts Ships Harmony Coding LLM Tuned for ArkTS · ByteDance's Doubao App Lands on Huawei HarmonyOS PCs With Doubao-Seed-2.1 Pro and Turbo

Simplexity Robotics, a full-stack embodied AI company, says it has trained a single robot to tend two CNC lathes on its own, loading parts into chucks with about 0.5 millimeters of clearance and unloading them after machining. Feng Zongbao, the company's head of reinforcement learning, presented the work in a Tech Talk at IROS 2026 in Pittsburgh on September 29, and Leiphone published an edited transcript on October 8. Training used 600 real-robot trajectories.

The company frames the task as robots making robots, since joints, chucks and precision parts come off CNC machines. Feng argued it exposes the gap between lab benchmarks and factory precision: the tight clearance leaves almost no room for alignment error, reflective metal and weak textures confuse vision, similar-looking parts are easy to mix up, and key contact surfaces are often hidden from cameras in a cramped cell.

Three components handle action generation, memory and action selection. SimpleWAM, the company's world action model, learns during training to reconstruct future video frames, 3D states and action chunks, but at inference takes the current images, 3D state and instruction and outputs actions directly. Head and wrist camera features are fused adaptively so neither view overwhelms the other. In nominal trials, the fused approach succeeded 15 times, against 8 for a π0.5 baseline trained with supervised fine-tuning.

DRAM, a Delta-Rule Recurrent Associative Memory, attaches to a frozen pretrained policy and carries context from earlier steps in a fixed-size memory matrix, so its footprint does not grow over a long task. DPE scores candidate action chunks with a separately trained evaluator while the action proposer stays frozen; successful and failed rollouts update only the evaluator, which the company says avoids policy drift while raising success rates and cutting execution steps.

Because vision alone cannot show how a part is seating, force and torque readings feed both the policy's state input and a real-time impedance controller, which supports recovery when a straight insertion is blocked. Simplexity reported a 100% task success rate on the precision CNC insertion task. That figure applies to the insertion task, not to the yield of the whole cell, and the talk did not give the number of trials behind it.

The demo, shown at five times speed, covered grasping a workpiece, loading the chuck, and removing and placing the finished part. The task ran on the company's i7 Pro wheeled robot, one of four hardware lines alongside its SimpleClaw developer platform.

SEE ALSO: Simplexity Robotics: 11-Month-Old Embodied AI Startup Ships 100 Robots to Production Lines · Galbot and Tsinghua's LATENT Wins an IROS Award for a Humanoid Tennis Forehand

Inference KV cache has grown large enough to need its own storage. Huawei's OceanStor M900, unveiled at Huawei Connect, and Nvidia's CMX Context Memory Storage both move reusable KV cache into a separate, shared tier. Yet the solid-state drives meant to hold that data still have no agreed specification, according to industry sources interviewed by Leiphone in a report published October 8. The outlet identified its sources only by pseudonyms.

The logic is to trade storage for compute. A storage researcher told the outlet that raising the KV cache hit rate from 90% to 95% cuts the share that must be recomputed from 10% to 5%, roughly halving that compute cost. The catch is that the last few points of hit rate require keeping far more historical KV data, and its useful life varies widely: some consumer apps clear it within minutes or hours, while enterprise workloads that rely on long-running context gain more from keeping it.

Where it lives is the second question. An executive at a leading storage vendor said DRAM would be the obvious home if it were cheap and plentiful. It is neither, which leaves room for SSDs, though the swap is not one-for-one. Replacing DRAM with SSD capacity tends to run at ratios of 1:N, sometimes 1:5 or 1:10, while cost per unit of capacity can differ by dozens of times. The researcher added that most inference deployments do not yet need a dedicated KV appliance and will first use existing DRAM, memory pools and local SSDs.

The missing standard shows up in endurance. The executive said one KV cache design he had seen first asked for about 3 drive writes per day (DWPD), then raised the requirement to 9, and "it could be 12 tomorrow and 5 the day after." Short-lived cache means heavy rewrites and high endurance needs; data kept for weeks shifts the pressure to capacity. With lifecycles ranging from minutes to weeks, most buyers still purchase ordinary enterprise SSDs.

Because SSDs depend on scale, a product category needs a common specification. If each large internet company defines its own drive, vendors cannot reach volume. The executive estimated it will take 6 to 12 months for the industry to converge, and roughly another year after that for development, validation and tuning.

Offloading also leaves other costs in place. Moving KV data to SSDs mainly eases DRAM and affects time to first token more than accelerator memory pressure. The founder of an AI chip company said compute-in-memory designs target a different cost, moving model weights and active data, and do not replace bulk storage for historical KV cache.

SEE ALSO: Huawei Debuts OceanStor M900 Context Memory Storage: PB-Scale Shared KV Cache With ~60μs NPU-to-SSD Hop · Huawei Shows Experimental XMFS at LPC 2026 for Cross-Node Shared Memory

ByteDance Seed researchers have identified a periodic weak spot in language models that compress their key-value (KV) cache in fixed-size chunks, a design DeepSeek-V4 uses to cut the memory and attention cost of long contexts. In a paper titled "Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression," posted to arXiv on September 28, the team reports that the same information can be easy to retrieve at one position and hard at another, a gap that average benchmark scores can hide. IT Home reported the work on October 9.

Chunked compression groups consecutive tokens into windows and squeezes each window into fewer cache entries at a fixed stride. That gives every token a new coordinate, which the authors call its phase: its position relative to the window boundaries. Shifting an input by a few tokens can change which tokens are compressed together without changing their content.

The team tested base and post-trained versions of DeepSeek-V4-Flash and DeepSeek-V4-Pro, plus the post-trained DeepSeek-V4.1-Flash, on a needle-in-a-haystack retrieval task at 128K tokens, holding prompt length and absolute position constant. Accuracy in the base checkpoints differed by up to 40.2 percentage points depending on the phase of the target. Post-training raised accuracy and narrowed the gaps, but they stayed large for the V4 models. DeepSeek-V4.1-Flash, which uses a stride of two rather than four, showed the smallest gap, though the periodic pattern remained.

To test the cause, the researchers pretrained a family of small transformers from scratch on a Qwen3-0.6B backbone, changing only the KV compression design. Every compressed variant showed periodicity matching its stride, including one that simply averages tokens in each window, while full-attention baselines did not. Interventions on attention heads point to what the authors call phase specialization, with different components handling information at different phases. They note that these mechanistic findings are strongest in controlled settings and do not establish a universal cause in large models.

The practical point concerns evaluation rather than any single model. The paper argues that models using chunked KV compression should be measured across phases, for example by shifting the input a few tokens while keeping the queried information fixed, because high average accuracy can coexist with systematic positional failures.

All eight authors list ByteDance Seed as an affiliation. Three also list Princeton University, Stanford University or the University of California, Berkeley, and the paper says the core contributors did the work during internships at ByteDance Seed. The corresponding authors are Yiyuan Ma and Xin Dong.

SEE ALSO: DeepSeek-V4.1-Flash Ships Causal Encoder–Decoder MoE With 1M Context and Extreme KV Compression · Xiaomi Details HySparse2 for MiMo-V3: Two-Level KV Sharing Cuts 1M-Token Prefill FLOPs About 5× vs Hybrid SWA

KargoBot said on September 22 that it has started a scaled AI freight network in Qipanjing, a town in Ordos, Inner Mongolia. The network puts three kinds of vehicles on one system: trucks with human drivers, autonomous trucks that still have a cab, and transport robots with no cab at all.

The company has spent about five years in Inner Mongolia on platooned driving. Leiphone reports a fleet on the order of 100 driverless trucks, energy routes reaching Xinjiang, Gansu, Shaanxi and Shanxi, and cabless robots now in tests. Last year, with Erdos Group, it linked mines, plants and loading docks so that route-level economics turned positive on real jobs. That is the company's own account.

KargoBot says cabless robots could lift the economic gain per vehicle from 20 percent to more than 30 percent. That 30 percent figure is the company's target, not an audited result. Chief executive Wei Junqing sketches a conventional cost split he calls 33211: tolls about 30 percent, energy about 30 percent, the driver about 20 percent, and maintenance and depreciation about 10 percent each. On current runs, the company says platooning improves gross margin by about 10 to 18 percent versus manned haulage, after depreciation, maintenance and insurance.

One lead driver, called an AI navigator, can in theory head up to five follower trucks, a line more than 100 meters long. On public roads KargoBot usually pairs the lead with one unmanned truck. If a car cuts in, the follower honks, slows and opens a gap. A red light can split the pair: the lead waits, and the follower drives itself back into formation. The follower is not a copy of the lead. It receives processed road data plus throttle, brake and path, but it keeps its own judgment if that link fails. On weak-signal stretches in Inner Mongolia and Xinjiang, the trucks exchange data directly.

The network now covers six provinces, 30 customers, eight truck makers and about 30 partners, and a new route takes about half the old 8 to 12 months, the company says. Most work is still run by KargoBot or large logistics firms; mature routes could later admit fleets of a few trucks to a dozen. Some followers have gone three to six months with no one at the wheel. Dropping the cab, by the company's estimate, frees about 25 percent more cargo space, so a convoy can mix all three vehicle types. KargoBot's tests also put the energy saving of a two-truck convoy at about 5 to 10 percent.

SEE ALSO: Pony.ai and GAC Unveil Gen-4 Level-4 Electric Robotruck at IAA 2026 · Plus Delivers 100 Factory-Installed Autonomous Heavy Trucks · Didi's "KargoBot" Autonomous Freight Service Operates Independently Following 450 Million Yuan Investment

Tencent's WorkBuddy can now carry a WeChat mini program from a plain-language request through preview and a publish step inside the same agent, according to a hands-on report on QQ News on October 8. The account is a product test by a tech creator, not a Tencent press release, and it is about the mini-program workflow only.

After updating WorkBuddy, the author turned on a mini-program mode inside the code tools and described a small voice notebook: speech in, text out, notes sorted for later. The agent started the project from that description. When the build finished, it returned a card. Opening the card brought up a side panel with a preview of the mini program and cloud tools, including the database. Where the request did not say what should be stored, WorkBuddy stopped and asked.

Cloud pieces that usually sit outside a code generator were connected along the way, including login, data storage and files. For speech-to-text the author picked Tencent's wand-asr-v1 model. For sorting the notes, WorkBuddy chose GLM-5.3-Flash. The report says the agent includes a model catalog and can attach a model when the mini program needs one, which is what happened in this test.

Publishing has more than one path. A developer who already has a mini-program account can scan a code to bind it and push an experience build or a release build. Someone who only wants to try the idea can create a trial mini program instead. The author says a rough version of the notebook was usable within a few minutes, with notes showing up in the database, while still calling the details unfinished.

None of that replaces WeChat's own review or filing. Those steps stay outside the agent. Binding an account, or asking for a release, is not the same as WeChat approving the app. The test shows a shorter path from a sentence to a preview, cloud storage and a publish request. It does not show that every mini program goes live on its own.

SEE ALSO: Tencent Opens WorkBuddy Platform, Onboarding More Than 100 Partners · WorkBuddy Goes to Work at Shanghai Jahwa: 9 Departments, 4.5x Average Efficiency · Weixin Mini Programs Expand Globally as Tencent Touts 100+ Country Reach and Uber Partnership

A paper from Galbot, Tsinghua University and collaborators won the Best Entertainment and Amusement Paper Award at IROS 2026, whose main conference closed in Pittsburgh on September 30. The work, Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data, is nicknamed LATENT. Peking University and Shanghai AI Laboratory are also on the team. The award is for a tennis return, not for a robot that plays a match.

The clips are deliberately incomplete. Five amateur players spent about five hours in a capture area of roughly 3 by 5 meters, recording fragments of forehands, backhands and footwork. They are not a match, and they do not say where the next ball should land. The paper calls the data imperfect because motion capture does not map cleanly onto a robot body, especially at the wrist, and because the clips cover basic movements rather than tactics.

LATENT trains a motion tracker, then compresses it into a continuous latent action space, a dictionary of body movements. Human clips are a prior, not a script. The team lowered how much the tracker trusts the right wrist, so a noisy joint cannot dominate the swing. A constraint called the Latent Action Barrier uses Mahalanobis distance so the policy does not jump between unrelated motions.

The scores that follow are from the paper. In simulation, the full forehand policy succeeded on 96.52 percent of returns, with an average landing error of 1.32 meters. Comparison methods AMP, ASE and PULSE reached 41.32 percent, 63.47 percent and 71.85 percent. Removing the barrier dropped forehand success only to 93.12 percent, but a smoothness score rose from 25.61 to 37.64 and joint torque from 7.40 to 12.53, which the authors read as jerkier movement.

On a real Unitree G1, the tests still use motion capture, not the robot's own cameras. Leiphone, citing project materials, says the rig used more than 50 cameras over 19 by 15 meters. Across 20 consecutive rallies, the paper reports 90.90 percent success on the forehand, 77.78 percent on the backhand, 88.89 percent from the front court and 81.82 percent from the back. Success means the ball landed in the opponent's court. A simulated robot-versus-robot rally of 25 shots is not from the real court. The paper lists motion capture as a limit and names active vision as a next step. The result is a lab forehand return, not a full tennis match.

SEE ALSO: Galbot Prices ET1 Bipedal Humanoid From 79,000 Yuan, With Watch-and-Learn Skills on Upper Tiers · Galbot Launches LDA-1B World-Action Model, Open-Sources Framework · Unitree G1 Nails a Perfect Layup: HKUST Showcases the World’s First Real-World Basketball Robot Demo

Huawei engineers presented an experimental filesystem called XMFS at the Linux Plumbers Conference in Prague on October 5. IT Home reported the talk on October 8. XMFS, short for eXpress Memory FS, is aimed at Linux servers that use CXL 3.0 or Huawei's unified bus. The aim is to let applications reach shared memory on other nodes through ordinary POSIX file calls, without rewriting the software above them.

Conference materials describe XMFS as a kernel prototype. The question it explores is whether memory that spans servers can sit inside Linux as a storage layer, rather than something each application addresses through its own interface. Public designs set it apart from other CXL and shared-memory filesystems such as FAMFS and DAXFS, which do not put a standard POSIX interface at the center. In principle, existing programs could open, read and write that remote memory the way they already use files.

Huawei told the conference it has run early tests on metadata-heavy workloads and on sharing container images across nodes. The write-up treats those tests as a sign that the approach can work, not as a product benchmark. The team also listed problems still open: keeping filesystem metadata in sync across nodes, providing software cache coherence when the hardware does not, treating memory-semantic links and RDMA as one transport, and managing and tiering capacity that lives on more than one machine.

Code has started to appear, but only on openEuler. A version-8 patch set posted to the openEuler kernel mailing list in August touches 15 new or changed files, covering core filesystem logic, directories, files, inodes, journaling, garbage collection and URMA support. That is not a submission to the mainline Linux kernel. XMFS has not entered formal mainline review, and the conference materials call it a prototype for exploring the idea, not a finished filesystem ready to merge.

For now, XMFS is research on Huawei's interconnect software stack. It is not a shipped filesystem, and it is not in the Linux kernel.

SEE ALSO: Huawei Positions Lingqu UnifiedBus as Core of Agentic SuperPoD Cluster Architecture · Huawei Debuts OceanStor M900 Context Memory Storage: PB-Scale Shared KV Cache With ~60μs NPU-to-SSD Hop · Global Voices on openEuler: Bridging Open Source, AI, and International Collaboration

The world's largest-capacity flexible DC transformer is now in place in the heart of Guangzhou. Rated at 770 megavolt-amperes, the 220-kilovolt unit has been installed at the Tianhe Tangxia flexible DC back-to-back project in the city's Tianhe district, according to China Southern Power Grid's Guangzhou Power Supply Bureau and state broadcaster CCTV. CCTV says its capacity is enough to slow-charge about 100,000 electric vehicles at once.

The three-phase, water-cooled transformer is the second main transformer delivered to the site and the core piece of equipment in the project. Conventional transformers at the 220-kilovolt level usually range from 120 to 240 megavolt-amperes and rarely exceed 350, so a single unit here does the work of several. Developing it meant solving problems including leakage flux suppression, balancing insulation against temperature rise, low-noise operation in a dense urban area and digital-twin maintenance.

The project addresses a structural weakness in the city's grid. Guangzhou's core is served by three supply zones, Mumian, Zengcheng and Guangnan, which have had limited ability to support one another. Shifting power between them has meant opening multiple switches and waiting for flows to rebalance, which can be slow and risks outages. Flexible DC technology acts like a smart sluice and two-way pump between the zones: it can send surplus power to a zone in need within milliseconds and cut flows off instantly if they become excessive.

Because the site sits on some of the city's most expensive land, the team stacked the equipment vertically. The converter station stands 49 meters tall, making it the tallest onshore flexible DC converter station in China. Once complete, the project will add supply capacity for about 350,000 households, according to CCTV, and a September 28 project update says it will provide 900 megawatts of reliable support to Guangzhou's core area.

Noise was a particular concern in a residential district. The transformer emits a distinctive hum between 100 and 400 hertz, so engineers developed perforated sound-absorbing panels that line the surrounding walls. Sound waves entering the holes resonate and cancel out, cutting transformer noise by an estimated 80% to 90%.

Everything from the core components to the auxiliary equipment was made in China, CCTV said. More than 450 workers stayed on site in shifts through the National Day holiday, and the project is scheduled for trial energization by the end of the year. The site will later host a substation, a BeiDou base station and an energy storage station in what is planned as the country's largest multi-station complex.

SEE ALSO: China Activates First 220kV Quantum Substation in Hefei — a 18-Month, 85-Device Quantum Deployment on the State Grid · POWERCHINA Connects 1-GW Gonghe Solar PV and Molten-Salt Tower Project in Qinghai to the Grid · China Huadian Connects First 100-MW-Class Compressed CO2 Energy Storage Project in Xinjiang to Grid

A team at Tianjin University led by Jiang Zhongyi and He Guangwei has developed an all-organic mixed-matrix membrane that removes carbon dioxide from industrial gas far faster than conventional polymer membranes while staying easy to process. The work, titled "All-organic mixed-matrix membranes with single-crystal covalent organic frameworks for carbon capture," was published online in Nature Energy on October 7, with collaborators including EPFL in Switzerland and the University of Southern California.

Flue gas from power plants, cement works and steel mills contains a low concentration of carbon dioxide but enormous total volumes, and conventional separation methods are energy-hungry and expensive. Membranes are a cleaner option, but industrial use demands high permeability, high selectivity, long-term stability and easy scale-up at the same time. Mixed-matrix membranes combine a processable polymer with a porous filler, yet the two materials rarely mix well. The usual fix is to make them bind more tightly, but performance then remains dominated by the polymer and the filler contributes little.

Inspired by the structure of cell membranes, the team took the opposite route with what it calls a weak-interaction inlaying strategy. By balancing the forces between filler and polymer against those between filler particles, they embedded single-crystal covalent organic frameworks, crystalline materials with regular, ordered pores, evenly into the polymer at loadings of up to 75.5% by volume. At that level the framework crystals become the main transport pathway for gas molecules, while the membrane keeps its flexibility and can still be cast from solution.

In tests with simulated flue gas mixtures, the membranes reached a carbon dioxide permeability of 74,800 Barrer with a carbon dioxide to nitrogen selectivity of about 20, beating previous records and reaching roughly 70% of the value predicted for a pure framework membrane. Commercial polymer membranes for post-combustion capture typically sit in the tens to low hundreds of Barrer. Performance held steady in the presence of water vapor, sulfur oxides and nitrogen oxides, and the paper reports that polyimide-based versions kept about 85% of their initial permeability after six months of aging at room temperature.

A preliminary techno-economic analysis, based on an assumed process design and cost conditions, puts the capture cost at about $38 per ton of carbon dioxide. The figure is an early estimate rather than a measured result from an operating plant.

"Turning the challenge of industrial carbon reduction into a scientific question about advanced membrane materials is what we have long worked toward," said corresponding author Jiang Zhongyi. The team says the approach could support flue gas capture in power, cement and steel.

SEE ALSO: China Huadian Connects First 100-MW-Class Compressed CO2 Energy Storage Project in Xinjiang to Grid · China's First Offshore Carbon Storage Project Achieves Over 100 Million Cubic Meters of Carbon Sequestration · Tianjin University Lab Unveils 3-Gram Non-Invasive Brain-Computer Interface That Hides in Hair

Xiaomi said on October 7, via the official HyperOS account on Weibo, that HyperOS 4 is now fully compatible with the Apple ecosystem. The update pulls together a set of cross-platform features that let Xiaomi phones work with iPhones, iPads and Macs with less friction.

Xiaomi highlighted four capabilities. Touch to share lets users tap devices together to send photos, files, contacts and browser pages across ecosystems. Users can view both the iPhone's and the Xiaomi phone's cloud photo albums from either side, covering Xiaomi Cloud and iCloud. Notifications sync in both directions, so important messages can be read and answered without switching phones. And Home screen+ now supports all three of Apple's main devices.

To use Home screen+, a Mac or iPad needs the Xiaomi Interconnectivity app from the App Store, signed in to the same Xiaomi account as the phone, on the same network or hotspot and with Bluetooth on. On the phone, users switch on device interconnection and Home screen+ in settings. Macs need macOS 12 or later and iPads need iPadOS 14 or later. Once connected, users can see and operate the Xiaomi phone from the Mac or iPad and drag images or files directly between apps such as the gallery and file manager.

Separately, the Xiaomi 18 Pro Max running HyperOS 4 has gained a transfer with Apple devices feature that works with AirDrop, according to screenshots shared by a Chinese tech blogger. The function relies on Google's Quick Share, which Google first made interoperable with AirDrop on the Pixel 10 series in November 2025 before extending it to more Android phones. The receiving Apple device must set AirDrop to Everyone for 10 Minutes, and Google says transfers travel over a direct peer-to-peer connection without passing through a server. Samsung, Pixel, the OPPO Find X9 Ultra and the vivo X300 Ultra already support it, and the Xiaomi 17T Pro added it in June 2026.

The work started earlier. HyperOS 3 already synced WeChat, QQ, text and call notifications with iPhones and let users unlock a Xiaomi phone with Touch ID, Face ID or a passcode on Apple devices. Touch to share requires an iPhone 12 or newer on iOS 17 or later. In September, Xiaomi also put its interconnectivity app on the Apple Watch App Store for notification forwarding, health data sync and ring-to-find.

HyperOS 4 debuted with the Xiaomi 18 Pro series on September 23, with stable updates rolling out from October 15. Xiaomi has not said when these Apple features will reach international versions.

SEE ALSO: Xiaomi HyperOS 4 Stable Rollout Starts October 15 With Xiaomi 17 and REDMI K100 Pro · Huawei & Apple Achieve Full Cross-Device Connectivity: HarmonyOS CrosShare for Mac Now Available on the App Store · September Phone Wars: Folding iPhones, Wide Foldables and a Wave of Price Hikes

UBTECH Robotics announced on October 8 that it has reached a strategic cooperation agreement with FAW-Volkswagen. The two companies will jointly develop and test application scenarios for embodied AI robots in logistics and build demonstration sites, with the aim of speeding up the deployment of humanoid robots in smart manufacturing and promoting humanoid smart-manufacturing solutions across the industry.

The deal builds on a deployment that is already running. UBTECH's Walker S Lite humanoid has been carrying out vehicle quality-inspection training at FAW-Volkswagen's Qingdao Branch, which operates a national-level smart manufacturing demonstration factory. According to earlier reports, the robot has been integrated with the plant's automation control system and handles complex tasks there with precision. The new agreement extends that work from inspection into logistics.

The partnership dates back to July 2024, when FAW-Volkswagen opened production line scenarios at the Qingdao factory to UBTECH's industrial humanoid Walker S for tasks such as bolt tightening, part installation and part handling. At the time, the two sides also said they would explore industrial logistics robots and supply chain cooperation. The latest agreement makes logistics a formal focus, with jointly built demonstration scenarios rather than one-off pilots.

Logistics is one of the most labor-intensive and repetitive parts of car production, involving the constant movement of parts and materials around the plant. That makes it a natural next step for humanoids that have already proven they can work alongside an existing production line and connect to its control systems.

UBTECH says its approach to real industrial deployment rests on four core technology clusters: an embodied brain, a humanoid cerebellum for motion control, high-performance robot bodies and swarm intelligence that lets multiple robots coordinate. For the embodied brain, the company has built a fully in-house stack spanning a foundation model, a world model and an action model, designed to let robots understand their surroundings, predict how they will change and decide what to do.

Production capacity is also ramping up. UBTECH's industrial humanoid super smart factory in Liuzhou, Guangxi, began operating on September 12. According to the Guangxi regional government, the 14,000-square-meter plant builds the Walker S and Cruzr series, can roll out one robot every 10 minutes and uses robots to help build robots.

With FAW-Volkswagen now extending humanoid work from inspection to logistics, the Qingdao plant is becoming one of the clearest tests of how many jobs a single humanoid platform can cover inside a real car factory.

SEE ALSO: UBTECH Commissions Liuzhou Industrial Humanoid Robot Super Smart Factory · UBTECH Partners with Honda Trading to Advance Commercialization of Humanoid Robots and Autonomous Vehicles · UBTECH Begins Mass Deliveries of UWorld U1 Companion Humanoids

ByteDance's AI assistant Doubao is now available as a native app on Huawei's HarmonyOS PCs. Chinese tech outlets first reported the desktop client's arrival on October 1, and IT Home confirmed on October 7 that users can download it directly by searching for Doubao in the HarmonyOS app store. Huawei said Doubao ranked first on the HarmonyOS PC user wishlist, the list of apps owners of its HarmonyOS computers most want to see on the platform.

The first version ships with two models, Doubao-Seed-2.1 Pro and Doubao-Seed-2.1 Turbo. ByteDance released the Seed 2.1 series on Volcengine in June, positioning Pro for exploring highly complex tasks and Turbo for large-scale production use. Compared with the previous generation, the company says Seed 2.1 improves coding delivery, long-chain agent task execution and multimodal understanding.

The app is built around two modes that users can switch between freely: Chat and Work. In Work mode, users set a goal and attach files, images or material from their Doubao cloud drive, then choose Auto, Turbo or Pro and pick a reasoning intensity. They can also create projects, add local folders and set scheduled tasks. A side workbench supports new tabs, so a user can browse the web, read reference material and talk to Doubao at the same time while several tasks run in parallel.

For content creation, the launch version includes deep research, PowerPoint generation, a writing assistant, image generation and cloud storage. Video generation, AI blogging, recording transcription, music generation and AI spreadsheets are also available. ByteDance pitches the client at office work, academic research, file management, automated tasks and creative design.

HarmonyOS PCs run Huawei's own operating system rather than Windows, which means every mainstream application has to be rebuilt natively before users can install it. A general-purpose AI assistant has been one of the most visible gaps, and Doubao's arrival gives Huawei's laptop owners access to one of China's most widely used chatbots without relying on a browser.

For ByteDance, the release extends Doubao's desktop presence beyond Windows and Mac, where the company has steadily turned its PC client into an agent that can operate software, run multi-step tasks and work across local and cloud files.

SEE ALSO: ByteDance's Doubao PC Client Has Quietly Become a Codex-Style Agent — GUI Control, Remote Tasks, and Cloud Computers · Doubao Work Adds Parallel Multi-Agents and Mac Local GUI Control · HarmonyOS Is Now the World's Third-Largest Mobile Operating System, Huawei Says

China's homegrown satellite internet has passed its first test in the polar regions. During the recent Mid-Autumn Festival, satellite communication equipment running on the GW constellation operated by China Satellite Network Group (China SatNet) enabled a video call between the Arctic and Antarctica, with a delay of no more than one second, state broadcaster CCTV reported on October 6.

In the test, research vessels taking part in China's 16th Arctic Ocean expedition connected by video with Qinling Station, the country's newest research station in Antarctica. Maximum uplink and downlink speeds reached about 30 Mbps. After the ships moved into ice-covered waters, the link was also used to transfer files and send navigation logs back to shore.

The GW constellation is China's first megaconstellation for satellite internet. It is planned to include thousands of low Earth orbit satellites, many of them in near-polar orbits that can provide coverage at high latitudes, where conventional geostationary satellites sit low on the horizon or drop out of view entirely. That makes the polar regions a demanding proving ground for both satellites and terminals.

According to the report, the equipment can also be used for emergency rescue and maritime communications. Reliable, low-latency links matter for ships working far from shore, for coordinating logistics and safety at remote research stations, and for sending scientific data home without waiting for a port call.

The expedition itself set out from Dalian, Liaoning, on July 3. Organized by the Ministry of Natural Resources, it involved four vessels, including the icebreakers Xuelong and Xuelong 2, and was the first Arctic Ocean expedition of China's 15th Five-Year Plan period. Xuelong and Xuelong 2 returned to Shanghai on September 30 after completing integrated surveys, frontier research and international cooperation tasks, with the expedition's scale and workload both setting new records. The two remaining vessels are still at sea and are expected to return in mid-October.

For polar science, where expeditions generate large volumes of ocean, ice and atmospheric data, faster satellite connectivity could shorten the gap between collecting observations and analyzing them, and make life aboard remote ships and stations a little less isolated.

SEE ALSO: China's Satellite-Making Sprint: Inside Wenchang's 1,000-a-Year Super Factory · Hainan Commercial Launch Site Phase II Enters Full Rehearsal After 612 Days, Adding Pads 3 and 4 · LandSpace's Zhuque-3 Y-2 Sets August 19 Launch Window