SAIC-GM-Wuling announced a smart equipment strategy on October 29 and launched its first product under it: an automotive-grade unmanned logistics vehicle for last-mile delivery. The move takes the carmaker, best known for low-cost cars and minivans, into driverless delivery.

The vehicle is built on what the company calls an architectural chassis, with drive-by-wire braking, steering and propulsion. Wuling says this gives it passenger-car-level dynamic safety. The cargo body is modular and can be fitted as an enclosed container, a fenced cargo bed or a refrigerated box, and the cab and cargo box are separated so goods can be loaded by forklift.

Published specifications list 6 cubic meters of cargo space, a combined range of 200 kilometers and a gross weight of 1,850 kilograms. The vehicle uses a 16.2kWh lithium iron phosphate battery that can charge from 30 to 80 percent in about 0.58 hours. Top speed is 60 kilometers per hour, maximum gradient 18 percent, turning radius 4.7 meters and wading depth 300 millimeters. It is rated IP55 and operates from minus 20 to 50 degrees Celsius.

The driving system was developed in-house. It does not need high-definition maps, relying instead on lightweight mapping, an end-to-end large model trained with the company's passenger-car driving data, and routes deployed from the cloud, so a vehicle can be set up quickly and sent on multi-stop deliveries with a single command. Wuling says it can run on open city roads where permitted as well as on closed industrial parks and campuses, for courier delivery, retail district delivery, cold-chain transport and in-plant material moves.

Wuling is leaning on its scale as a carmaker. The vehicle shares its battery, electric drive and supply chain with the company's passenger models and is built on automotive production lines that run both standard volume production and small custom batches. The company points to a national service network of more than 2,800 outlets and 15,000 technicians, with round-the-clock support and remote monitoring for fleet upkeep.

The company said the vehicles are already in operation with several partners, covering operators, technology providers and end customers. It signed a framework agreement with a smart-mobility unit of China Merchants Group, under which Wuling supplies vehicles, warranty and battery and drivetrain maintenance, while the partner handles local regulatory access and operations. A national standard for road testing and demonstration of intelligent connected vehicles, which took effect on July 1, set out unified road rights and responsibilities for unmanned delivery vehicles.

SynDexy, a Beijing embodied AI company focused on high-precision industrial manipulation, said on October 30 that it has completed the first small-batch delivery of its SynDexy C1 dexterous humanoid robot. The company says the product is moving from research and validation into engineering delivery and real deployments.

The robots will be used for precision assembly steps at customer factories in consumer electronics and electronic components. These steps demand high accuracy and consistency. In earlier tests, SynDexy says, C1 achieved a success rate above 98 percent on key tasks.

C1 is a two-armed, wheeled humanoid, and its main feature is the combination of vision and touch. On the hardware side it carries in-house visuotactile sensors that the company says offer micron-level spatial resolution and highly dynamic fine sensing, giving the robot real-time feedback for dexterous grasping and compliant handling. On the software side it runs SynDexy M1, a manipulation model that fuses vision, touch and force. SynDexy calls M1 the first model of its kind, a claim it has not backed with outside comparisons. The robot has 22 degrees of freedom across its body and runs motion control at a 1-millisecond cycle.

The development timeline was short. Work on C1 started in April 2026, the first prototype was finished in July and the first small batch shipped in October, about six months in total. The company attributes the pace to defining the product around real production line needs from the start.

SynDexy was founded in 2025. It builds its own robot hardware, visuotactile sensing and manipulation models, and targets precision manufacturing such as consumer electronics, semiconductor micro-components and auto parts. Its chief scientist is Fang Bin, a professor at the School of Artificial Intelligence at Beijing University of Posts and Telecommunications, who has worked on embodied perception, interaction and dexterous manipulation for more than a decade. The core team comes from Tsinghua University, Beihang University, Harbin Institute of Technology and Beijing University of Posts and Telecommunications, as well as Huawei, DJI and Tencent. Researchers make up more than 90 percent of staff.

Fine assembly in electronics plants, such as handling small connectors and flexible parts, remains one of the hardest jobs for robots because parts vary and the margin for error is small. Touch sensing is one of the main ways companies are trying to close that gap.

The company says it will keep focusing on high-precision dexterous manipulation in industry. It did not name the customers that received the first robots or say how many units were delivered.

Midea's six-arm humanoid robot MIRO U has been working for more than three months at the company's washing machine factory in Wuxi, where it has fastened the back-panel screws on 100,000 washing machines. The figure was given by the head of the KUKA embodied robotics innovation center during a media tour organized in mid to late October by the Guangdong Department of Industry and Information Technology, according to a report by Time Finance carried on QQ News.

At the Midea Global Innovation Center lab in Shunde, Foshan, reporters watched MIRO U drive screws at an average of six screws in about 10 seconds, with a reported task success rate of 99.8 percent.

Midea, which owns German robot maker KUKA, says the six arms are a practical choice. Two-armed humanoids are flexible but work at only around 30 percent of a person's efficiency, which is not enough for an appliance production line. Putting six arms on one machine lets its output per square meter of floor space catch up with skilled workers. Midea first showed MIRO U in December 2025 as the third generation of its humanoid family, a wheeled robot that can lift itself vertically and rotate 360 degrees on the spot. The company follows an industry-first path: robots are proven at its own washing machine plants in Jingzhou and Wuxi and a logistics warehouse in Wuhu before being offered to outside customers.

MIRO U is part of a wider push in the province. As of September, more than 2,000 humanoid robots had been placed in manufacturing companies in Guangdong this year and are running in real production settings, and the number is expected to pass 3,000 by year-end, according to the report.

Other examples from the tour show how humanoid work is shifting from demonstrations to measurable tasks. At X Square Robot's Shenzhen headquarters, a two-armed robot flips parcels so their labels face up, sorting more than 1,800 packages an hour. AI² Robotics said its robots work in semiconductor display, automotive and biopharmaceutical plants. Executives still described large-scale rollout as some way off. A Huayan Robotics executive said solutions that work well in simulation often lose accuracy in real workshops because of changing light, part variation, vibration, oil and cycle-time pressure.

Guangdong's official data shows the province produced 336,300 industrial robots in 2025, 43.5 percent of the national total, and that the Pearl River Delta can produce prototypes within 48 hours and deliver small batches within seven days.

Tencent has opened Tencent Beacon, its AI data analysis platform, to outside companies and individual users, who can now register at beacon.qq.com. For about a decade Beacon served Tencent's own teams for self-service analysis and data visualization. It has now been rebuilt as an AI data analysis suite that lets business staff ask questions in plain language, generate dashboards, trace the causes of metric changes, set up monitoring and work together, without writing SQL.

Tencent says the upgraded Beacon AI is used by more than 100 internal teams, including QQ, Tencent Video, Tencent News, WeChat Pay, Yuanbao and TiMi game studios. In examples cited by the company, an end-to-end operations dashboard that took four to six working days in traditional BI tools was delivered in under two hours, and a monitoring dashboard was finalized eight minutes after the user opened the chat window.

Beacon AI connects to three categories and 10 types of data sources. These include MySQL, StarRocks, PostgreSQL and ClickHouse databases, local Excel and CSV files, and online documents from Tencent Docs, WeCom smart sheets and Feishu. Users can tag fields as dimensions, metrics or time dimensions so the AI reads them correctly.

A user mentions a dataset in the chat box and describes what they want, and the AI streams the result as it understands the request, matches fields, picks charts and renders the dashboard. Dozens of chart types are supported. Dashboards can be edited by pointing at an element and describing the change, by adjusting the AI-generated SQL, or by editing the dashboard's HTML directly, with version history and rollback. For attribution, users click an unusual chart and ask why a number moved. Alerts can fire on thresholds, empty data or failed queries and be pushed via WeCom or email.

The platform ships with 12 data analysis skills and three visualization skills, supports more from a skill marketplace and team-built skill packs, and connects to internal knowledge bases and other systems through the Model Context Protocol, or MCP. A feature called Self-Harness turns feedback, corrections and successful methods from past tasks into reusable experience.

On data security, Tencent says the model only sees metadata such as field names and data types during reasoning. Actual values are filled back in through placeholders and are not read by the model, and every step keeps an audit log. Team spaces offer three permission levels, and administrators can limit which AI models members use. An enterprise edition supports private deployment.

Zhuoyu Technology, the driver-assistance company formerly known as DJI Automotive, launched a driver-assistance solution for commercial vehicles at an event in Zhuhai on October 29. It comes in two versions. PDS 4 runs on the company's end-to-end driving model version 4.0 and later, supports highway navigation on autopilot, or NOA, and targets heavy trucks on trunk logistics routes. It is already in mass production. PDS 5 will run on Zhuoyu's next-generation native multimodal foundation model and add urban NOA for buses, light trucks and vans.

Rather than build a separate system for trucks, Zhuoyu moved its passenger-car model across. Executives summed up the approach as a unified base, scenario enhancement and engineering adaptation. The company says its passenger-car system has gathered more than 3 billion kilometers of driving data. On top of that, it trained the truck version with about 400 million frames of heavy-truck data from trunk-route driving.

Trucks pose different problems. A fully loaded heavy truck can weigh around 40 tonnes, has large blind spots and needs far longer to stop. Zhuoyu says the truck braking chain has a delay of close to one second. Its in-cabin lidar system sits inside the cab rather than outside, which reduces exposure to rain, snow, mud and stones, and the company says it detects obstacles up to 500 meters away on highways, including small, low-reflectivity objects such as tires and cardboard boxes. The system also estimates trailer posture, the angle between tractor and trailer, and real-time load, and coordinates the engine, retarder, auxiliary brakes and service brakes on slopes and curves.

Zhuoyu also added a fuel-saving algorithm trained with reinforcement learning. It decides when to lift off the throttle, apply power or use the retarder, based on gradient, curves, the powertrain and surrounding traffic. Across more than 200 production trucks tracked over two months and more than 20 million kilometers, the company reports fuel savings of 4.5 to 7.5 percent.

The first production vehicle is the Shacman X5000 from Shaanxi Automobile Group, which entered production in June. Several hundred units now run on trunk routes for courier company Yunda. Zhuoyu says it is also working with XCMG, Sinotruk, FAW Jiefang, Foton and Dongfeng Commercial Vehicle, with several models due by the end of the year. Yutong Bus will use PDS 5 on a city coach.

The system remains driver assistance that requires continuous supervision, not driverless operation.

Researchers from Tsinghua University, Infinigence AI and AgiBot have proposed Collective Experience Learning, or CEL, a training approach in which robots learn from the successes of other robots rather than mainly from human demonstrations. The team validated it on a cluster of 96 real robots, according to a release published by QbitAI on October 30. The accompanying paper is listed on arXiv as 2610.25736.

The team's starting point is a staffing problem. Robots improve by gathering real-world experience, and today that usually means a person demonstrating a motion again and again or stepping in to correct mistakes. Robots can be deployed by the hundreds, but experienced operators cannot be scaled at the same rate. Under CEL, the successful experience one robot gathers is stored and shared, so a robot that struggles with a task can look up how its peers solved similar situations.

The researchers tested five tasks spanning home and factory work: arranging flowers, folding clothes, charging a phone, assembling a stool and packaging a phone. Two robots that had not taken part in training ran 600 trials and failed 15 times, an overall success rate of 97.5 percent. Success rates on all five tasks were higher than the average of six human operators. On phone packaging, a multi-step industrial task, success rose from 79.2 percent to 98.3 percent after collective learning, with two failures in 120 consecutive runs.

The team also reported a scaling trend. With the same amount of real training time, raising the number of participating robots from 16 to 96 lifted success from 86.7 percent to 98.3 percent.

Two components make the setup work. The first is a fault-tolerant asynchronous training system, called FATS, built on the open-source large-scale reinforcement learning framework RLinf. Robots upload their experience as soon as they finish a task, without waiting for others, and the cluster keeps running when devices drop offline. In the 96-robot test, the cluster collected 9,834 episodes. Twenty-two robots had 30 faults in total, but lost robot time amounted to only 1.3 percent of total operating time.

The second is an algorithm called failure-guided experience retrieval fine-tuning, or FERT. A single manipulation can last minutes and involve thousands of action steps, yet the only feedback is often a final success or failure. FERT uses both: when a robot runs into trouble, it retrieves successful experience from similar situations and focuses learning on those examples.

The release was supplied by Infinigence AI. The results are reported by the research team and have not been independently replicated.

SEE ALSO: AgiBot Unveils GO-3 Preview Robot Model, Tops Three Leaderboards and Plans Open Source

A noninvasive ultrasound brain-computer interface developed by teams including the Shenzhen Institute of Advanced Technology of the Chinese Academy of Sciences was shown this week at a brain-computer interface industry conference in Shenzhen, according to a report by CCTV News.

Unlike most brain-computer interfaces, the system does not require opening the skull or implanting electrodes. Instead of only reading signals from the brain, it sends precisely targeted ultrasound into it. Researchers describe this as a way to write regulating signals into brain circuits and potentially activate the brain's own ability to rewire, rebuilding damaged or inactive neural pathways.

The technology is being studied for several conditions. In a preliminary clinical study with Kunming Medical University, 13 of 24 participants with chronic disorders of consciousness, a group that includes patients often described as being in a vegetative state, showed clear improvements in conscious behavior. The protocol consisted of three 20-minute sessions a week for four weeks.

The report highlighted one patient who had been unresponsive for three months after a brain hemorrhage. After 30 days of ultrasound treatment, the patient regained consciousness and, following rehabilitation, can now hold conversations, eat independently and walk with assistance. Individual cases like this do not show how well the approach works across patients, and the study was small and early.

Researchers say the intervention can be tailored to different conditions. In epilepsy, where a brain region fires abnormally at high frequency, ultrasound is used to suppress the overexcited area. In depression, which involves imbalanced emotional regulation circuits, it is used to modulate those circuits. In disorders of consciousness, where communication between the thalamus and the cerebral cortex is impaired, repeated rounds of gentle ultrasound are intended to help the network gradually recover. The epilepsy and depression work is still at the intervention research stage.

Brain-computer interfaces are usually grouped into invasive, semi-invasive and noninvasive approaches. The researchers present focused ultrasound as a new path within the noninvasive category, which avoids surgical risk but has traditionally offered less precise access to deep brain structures.

The results reported so far come from a preliminary study described in media coverage. Larger controlled trials would be needed before the approach could be considered an established treatment, and no regulatory approval was mentioned.

Tencent Hunyuan has published a technical report for Hy-Nomas, its open interactive world generation model, and released Hy-Nomas-Bench, a benchmark for testing interactive video world models. The report went up on arXiv on October 27, following the initial code and weight release on October 21.

Hy-Nomas is a video world model with an 8B diffusion backbone. Users steer it over multiple turns with free-form text instructions, continuous camera control, or both. It generates the world segment by segment, conditioning each new segment on the current input and on a cached history so that scenes stay consistent over time.

The report says the central design aligns structured video captions with action-aware causal attention. Descriptive context stays available across segments, while action instructions only apply to the time window they are meant for. A vision-language prompt enhancer turns user requests and the interaction history into structured conditions for the next segment.

Training follows a multi-stage curriculum. It introduces causal generation first, then text interaction and camera control, then training on histories the model generated itself, before a final diffusion distillation step. The distilled model runs in four to eight denoising steps. In human evaluation, the team says Hy-Nomas outperformed existing world models.

Hy-Nomas-Bench contains 280 test cases across four tasks: multi-turn free-form text interaction, camera navigation, interleaved text and camera control, and spatial memory. It checks whether a model follows successive instructions, produces physically plausible responses, executes camera moves accurately, and keeps a scene consistent when the camera returns to a place it has seen before. The suite includes prompts, structured captions, interaction sequences and reference camera trajectories, and is released under Apache 2.0.

The model itself is also Apache 2.0 licensed. Tencent provides an INT8 configuration that runs on a single RTX 4090 using sliding-window history and CPU offload, a Diffusers pipeline, LoRA fine-tuning code and an interactive web demo. Released adapters include a cinematic LoRA and a robotics LoRA trained on real dual-arm robot data.

Outside developers have already added support in SGLang, ComfyUI and WanGP, and one community demo reports real-time 720p generation on Google TPUs.

SEE ALSO: Tencent Hunyuan Open-Sources Hy Memory 2.0, Adding Working and Code Memory for AI Agents

Huawei will not develop drugs or build its own drug pipelines, according to the president of its pharmaceutical business unit. Speaking to media at the 11th China Pharmaceutical Innovation and Investment Conference in Nanjing on October 24, he said Huawei instead wants to be the technology base for AI drug discovery, providing computing power, storage, networking, toolchains and data governance to drug makers, research institutes and AI drug discovery companies.

The unit is one of Huawei's so-called legions, teams the company set up to focus on specific industries. It began appearing at pharmaceutical industry forums in late 2025.

The executive described AI drug discovery as fast-moving but still carrying high expectations. Some technologies, such as protein structure prediction and molecular simulation, are already in real use, but the full drug development chain has not been connected. The next question, he said, is not how large models are, but whether they can enter real research workflows and form a closed loop with lab data and research decisions. Whether AI can raise efficiency and help companies produce a steady stream of products will be proven only by actual pipelines and results.

Some partner projects have started deploying clusters of about a hundred accelerator cards to train, deploy and optimize drug discovery models. Huawei has built such clusters together with two healthcare companies, and also supports research bodies including the Shanghai Institute of Materia Medica and Changping Laboratory.

He listed three advantages for Huawei. First is access to advanced computing and storage that many companies lack. Second is the ability to cluster that compute. Training a model needs at least 64 cards and often hundreds or thousands, which have to be linked into a cluster rather than used one by one. He pointed to Huawei's 1,024-card SuperPoD design, which he described as the largest of its kind in the world. Third is the ecosystem: Huawei works with DP Technology, StoneWise and MoleculeMind on joint solutions, with Huawei handling the coordination and adaptation of compute, storage and networking while partners build the models.

The position mirrors how Huawei approaches other industries, where it supplies infrastructure and leaves applications to partners. In drug discovery, that means betting on demand for large training clusters as more companies move from small experiments to production-scale models.

SEE ALSO: First Ascend 950DT 1,024-NPU SuperPoD Cluster Goes Live in Wuhan, Built in One Month

AgiBot launched GO-3 Preview on October 29, an embodied AI model that the company says can keep learning at scale across pre-training, post-training and after deployment. AgiBot plans to open-source the full model by the end of 2026.

GO-3 Preview is built on what AgiBot calls a WRAM, or World Reasoning Action Model, architecture. It links task reasoning, prediction of what will happen next and action generation, so a robot can understand an instruction, anticipate the outcome and produce motions it can actually carry out. The model has recently taken first place on three embodied AI leaderboards: RoboDojo, RoboTwin and RoboColiseum.

AgiBot says the model scales along two lines. On the data side, first-person human videos supply rich examples of objects being handled and tasks unfolding, while robot trajectories tie observations and actions to physical results. On the model side, stronger learning capacity turns this scattered experience into skills that transfer to new tasks.

After pre-training, the company says robots can complete unfamiliar tasks in homes, kitchens, multi-robot setups and dexterous manipulation without extra fine-tuning. With a few demonstrations, the model can learn long, delicate jobs such as making a jianbing, a Chinese savory crepe. After fine-tuning, and with an agent coordinating the steps, a robot can also assemble Lego pieces.

For post-training, AgiBot built Genie Evolver, a distributed training system that combines about a hundred real robots with physics simulation and world models. In tabletop curling, a robot surpassed the average tested human after four hours of online practice. In a block-tower game, it pulled out five blocks in a row while keeping the tower standing.

After deployment, the Genie Master agent framework connects judgment, execution and feedback so robots can keep improving on the job. An AgiBot G2 robot working in a real convenience store located the causes of its own failures, adjusted its approach and stored what it learned, without human demonstrations or tuning. Its success rate on the full task flow rose from 28.6 percent on the first day to a steady 85.7 percent after five days. On a task such as hanging small hooks, AgiBot says this method improved single-task success by 17.7 percentage points over conventional supervised fine-tuning, and it continues to learn online.

SEE ALSO: AgiBot's GO-3 Nano Preview Tops RoboDojo's Simulation Leaderboard With a 61.90 Average Score · AgiBot's Genie Master Lets G2 Robots Diagnose Their Own Failures on a Real Convenience-Store Shift

Doubao Work, ByteDance's AI office agent app, said on October 29 that it is opening third-party models to business customers. The first batch includes Kimi K3.1 Pro from Moonshot AI, GLM-5.3 from Zhipu and DeepSeek-V4.1-Flash, which appear in the app's model picker next to ByteDance's own Doubao models.

The update applies to the team and enterprise subscription plans and is switched on by default. Companies do not need to buy access from each model provider or file a separate application. Usage is deducted from the AI credits the organization already holds in Doubao Work.

Administrators get control over who can use which model. They can grant access to individual members, departments, groups or the whole company. They can also enable every model from one provider at once, or set up a single model on its own.

The move changes how Doubao Work positions itself. Until now, the app was largely a showcase for ByteDance's Doubao models inside an agent that can work across files, browsers and desktop apps. Letting enterprises route tasks to rival Chinese models makes the agent layer, rather than the underlying model, the main product. It also lets teams choose a model by task, for example a coding-focused model for engineering work and a different model for long documents.

It also fits a wider pattern among Chinese desktop agent products. Xiaomi's MiMo Desktop added DeepSeek, GLM, Kimi and Qwen models this week, and several other agent tools have made similar moves. For enterprise buyers, a single bill and one permission system across several model makers removes much of the procurement work of trying them side by side.

Doubao Work has been adding features at a steady pace. In recent weeks it added the ability to save recorded computer actions as reusable skills, one-click import of chat history, skills and memory from other AI tools, and an infinite creation canvas. It also supports parallel agents and local GUI control on Mac.

ByteDance did not say which other third-party models would follow the first batch, describing the current lineup as the first group of models to join.

SEE ALSO: Doubao Work and Tencent WorkBuddy Both Push Desktop Agents Further Onto the Phone

DeepSeek Harness, DeepSeek's open-source agent runtime, published its v0.2.2-alpha.2 pre-release on GitHub on October 28. The headline addition is an experimental plugin called subagent_oar, which lets the main agent hand a self-contained task to a Codex CLI or Claude Code installation on the same machine.

Harness has supported Claude Code and Codex before. The v0.1.0-rc.8 update in August let users install them as profile bundles inside its subagent system. The new tool works differently. It gives the main agent one entry point for dispatching work: when it calls subagent_oar, it picks either codex or claude, then submits a task description, a full prompt and an optional working directory. The tool returns a task ID right away, and when the job ends, a completion notice brings the final answer or the error back into the main session.

The external tools run as themselves. Codex CLI and Claude Code keep their own sign-in, models, tools and settings, so users need to install and log in to each one and make sure the Harness host can find it. Provider settings can pass a model, reasoning effort, launch arguments, environment variables and a timeout, depending on what each runtime supports.

The documentation sets clear limits. The main conversation is not forwarded, so the prompt has to spell out which files matter and what should come back. Tasks are one-shot, with no follow-up questions, session resume or live progress stream. Only the final answer or failure returns, not the outside agent's internal reasoning or tool history. The official examples enable a writable workspace and turn off permission prompts, and cancelling a task does not undo file changes the outside agent has already made.

Other additions help users check what delegated work did. An experimental change-request feature can show the status and diff of a linked pull request for a session running in its own Git worktree. A developer statistics plugin breaks down context, cache use, and model and tool time by turn and step. Small background-task outputs now arrive inside the completion notice, and subagent wait guidance cuts down on repeated polling.

The release also lets users update third-party plugin packages, rolling back if a hot update fails, and adds Japanese, Spanish, French, German, Russian and Brazilian Portuguese interface text. Earlier community pieces, including Bilibili creators testing multi-agent setups, came from users and are not DeepSeek products.

SEE ALSO: DeepSeek Harness v0.2.1-rc1 Arrives as an Unannounced V4.1 Flash Snapshot Briefly Surfaces · DeepSeek Harness Ships First Major Post-Beta Update: 14 Changes Bring Multimodal and Subagent Upgrades

XHEART, the automotive chip company spun out of Momenta's chip program, launched its X7 Ultra chip on October 29. The company describes it as a chip designed around large driving models, and rates it at 700 TOPS of compute for large-model workloads.

XHEART's chief executive said the design starts from a simple observation: driver-assistance software is moving from many small models toward a single end-to-end large model, so the chip has to change with it. When the company was founded in 2023, it decided to work backward from the models actually used in production cars, reorganizing compute, memory and data movement around them and leaving headroom for future models.

The X7 Ultra uses a very large NPU core. XHEART says moving data inside one core is orders of magnitude faster, at the terabyte-per-second level, than moving it between cores. A multi-level cache built for high capacity and high concurrency is meant to cut inference latency and power draw. A dedicated operator engine accelerates key Transformer operations such as LayerNorm, Softmax and Transpose, and supports FP16 and FP32 vector math so that models keep their precision.

According to the company, production models running on the chip match their reference outputs at 99.99 percent, and compute utilization reaches 98.7 percent. Compared with what it calls industry flagship products, XHEART claims an 80 percent cut in memory bandwidth consumption and double the energy efficiency. These figures come from the company and have not been independently tested.

The chip also carries what XHEART says is China's first mass-produced Armv9 automotive-grade CPU, with 20 percent higher performance and 10 percent better energy efficiency. A large-cluster layout improves multi-core coordination and L3 memory access, and a new SVE2 vector engine offers stronger signal-processing performance than the previous NEON generation.

XHEART also gave a first look at its next chip. The X9 targets 3,000 TOPS of large-model compute and 700GB/s of memory bandwidth per chip, and is planned for 2027. It will be designed together with next-generation world models, raising compute and memory access in step so that larger models can run efficiently.

The company says it follows a cadence of producing one generation, developing the next and researching a third. It took two years from founding to mass production. The first X7 chip was powered on within 24 hours of returning from the fab, and on day 17 the chip, low-level software and driving model ran together in a complete vehicle.

TRAE, ByteDance's AI development product, has launched a personal assistant for developers, which it calls TRAE Bao. The feature went live on October 29. TRAE announced it through its official channels and on its official Chinese community forum. Instead of handling one task per chat session, the assistant is designed to keep track of a developer's projects, tasks and decisions over time.

TRAE frames the problem this way: much of the time developers spend with AI coding tools goes not into writing code but into breaking down projects, assigning tasks and switching between sessions. Every new task means explaining the project background, tech stack and earlier decisions again. Complex requirements must be split up by hand, tracked across sessions and stitched back together.

The assistant uses a two-layer design. A global assistant handles memory and coordination across projects. It can read a developer's past tasks, conversations and outputs and pull out reusable solutions and constraints. When a developer starts something new and states the goal, the global assistant looks for relevant experience in earlier projects and loads the context automatically.

A project assistant focuses on one project at a time. Working from that project's context, it breaks a complex requirement into several executable tasks, runs subtasks in parallel, integrates them and summarizes the results. In TRAE's description, the developer's role shifts toward hearing reports and making decisions.

In the TRAE desktop app, the global assistant is reached through a dedicated entry point, while creating or opening a project leads to the project-level assistant. The global assistant can be linked to messaging apps such as Feishu and WeChat, so developers can send it tasks and receive progress updates without opening the client.

There are limits. The current version supports local tasks only. TRAE says cloud tasks and mobile support will arrive in later updates. Users on the official forum report that the feature is rolling out gradually, that it has appeared in the China version first, and that it currently uses automatic model selection without a manual model choice.

The launch adds TRAE to a growing group of desktop agents that let users hand off work from a phone or chat app and keep long-term memory across sessions.

SEE ALSO: Doubao Work and Tencent WorkBuddy Both Push Desktop Agents Further Onto the Phone

Tencent used its 2026 Global Digital Ecosystem Summit on October 29 to launch Buddy One, a system that ties its productivity apps for individual users to a single AI agent. Alongside it, Tencent released a new cloud storage product, Tencent Drive, designed to act as the storage layer for that agent setup.

Buddy One links Tencent Docs, Tencent Meeting, Tencent Drive and the knowledge tool ima, all running on the WorkBuddy agent. Users get one account across the apps and one shared pool of AI credits. Tencent says memory sharing between the apps will follow, and more products will join over time.

In practice, users link their accounts from WorkBuddy's PC client, version 5.9.0 or later. After that, they no longer need to tell the agent which app to search or paste context in by hand. A request triggers a cross-app search over documents, meeting records and knowledge bases the user has authorized, and the answer labels which app each source came from. Credits bought or earned in WorkBuddy become Buddy One credits that work across WorkBuddy, Tencent Docs, ima and Tencent Drive. A Buddy One VIP plan costs 700 yuan per month and bundles WorkBuddy's top tier, Tencent Meeting Pro, Tencent Docs premium membership, shared credits and 1TB of storage.

Tencent Drive is the more developer-relevant piece. Tencent describes it as cloud storage built for agents. Besides serving as the shared store for WorkBuddy, Tencent Docs and other Tencent apps, it can connect to supported third-party agents and applications through a command-line interface and the Model Context Protocol. That lets a user finish a draft with one AI tool and pick it up in another without re-uploading files. The WorkBuddy knowledge library feature already runs on it.

For businesses, Tencent Drive runs on Tencent Cloud Object Storage, with end-to-end encryption, off-site disaster recovery backup and VPC deployment. Admins can set fine-grained permissions for both employees and AI tools. Tencent also raised storage in WorkBuddy's enterprise edition from 5GB to 50GB per seat.

Tencent also said WorkBuddy will open its architecture further, letting companies plug in their own agents, third-party agents and third-party messaging systems. Tencent's cloud chief Dowson Tong said the choice of model, agent and messaging app should rest with the enterprise.

SEE ALSO: Doubao Work and Tencent WorkBuddy Both Push Desktop Agents Further Onto the Phone

DiffuSpace, the team behind the Dream 7B diffusion language model, has released and open-sourced Clover-35B-A3B-Preview. It is a diffusion language model aimed at the kind of work usually handled by conventional large language models: writing code, calling tools, browsing web pages and generating front-end interfaces.

Most language models are autoregressive. They predict one token at a time, each new token depending on the ones before it. Diffusion language models work differently. They update many positions at once and refine the output over several rounds, recovering masked or noisy text into a coherent answer. The approach promises faster generation, but diffusion models have historically been unreliable at agent tasks, with malformed tool calls and repeated failed steps.

Clover uses block diffusion. Output is split into blocks that are generated in sequence, while positions inside each block are denoised in parallel. The team says it restructured a Gated DeltaNet backbone for this block-level parallel denoising and added learnable block-size calibration. The model is a mixture-of-experts design with 35 billion total parameters and about 3 billion active per token, and it supports a thinking mode.

According to figures published by the team, Clover scores 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. On tool use, it reports 65.3 on the tau-cubed overall benchmark and 67.9 on BFCL-v4. The team says these results are ahead of autoregressive models of the same size, including Qwen3.6-35B-A3B. All of these numbers are self-reported and have not been independently verified.

Speed is the main selling point. DiffuSpace reports 515 tokens per second on agent tasks on a single NVIDIA H200, compared with 365 tokens per second for a DFlash2 speculative-decoding setup, and 680 tokens per second overall on a single GB300. In a web-agent test across 49 tasks, the team says Clover's median execution time was 1.52 seconds.

The weights are available on Hugging Face, GitHub and ModelScope under the Apache 2.0 license. DiffuSpace is also offering a hosted trial through its website and says an agent harness and a desktop coding client are in development.

Clover is labeled a preview, and its real-world reliability on long agent workflows will need outside testing. Still, it is one of the largest open diffusion language models aimed squarely at coding and agent use.

The first computing cluster built on Huawei's Ascend 950DT 1,024-card SuperPoD has been switched on in Wuhan. It was lit on October 28 at a Huawei artificial intelligence conference for Hubei province, held as part of Huawei's 2026 China roadshow.

The cluster was built with investment from Hubei Science and Technology Investment Group, a provincial state-backed investor. According to the group, the project took only one month from kickoff to power-on, a short timeline for a system of this size.

The deployment is a single liquid-cooled SuperPoD. One node carries 1,024 Ascend NPU accelerator cards. Both the compute cabinets and the interconnect cabinets are fully liquid-cooled. The system delivers 1 EFLOPS of FP8 compute and has 256TB of globally unified memory. Card-to-card bandwidth is at the terabyte-per-second level, and end-to-end communication latency is in the microseconds.

The design relies on UnifiedBus, Huawei's interconnect architecture for SuperPoDs. It lets every card in the cluster access the others directly, so the 1,024 NPUs can behave much like one large accelerator rather than many separate servers stitched together over a conventional network. Huawei positions the setup for training models with trillions of parameters and for running AI agent inference.

That direct-access design is the main point of the SuperPoD approach. Large model training spends a growing share of its time moving data between chips. Keeping memory unified across the pod and the interconnect inside the pod reduces how much traffic has to go through slower links.

The cluster will serve local demand. It is meant to give model developers and companies in Hubei a nearby source of compute for training large models and for building, tuning and running industry AI agents. Local reports describe it as filling a gap in high-end computing supply in the province.

The Wuhan launch is also an early real-world deployment of the Ascend 950 generation. Huawei first showed a physical Ascend 950 SuperPoD earlier this year and said at the time that more than 750 sets of its previous-generation Ascend 384 SuperPoD had already been deployed commercially. The 950DT cluster in Wuhan shows the newer SuperPoD design moving from showcase hardware into regional AI computing centers.

Huawei has not published the cluster's power draw or detailed utilization figures, and the performance numbers above are those given at the launch event.

SEE ALSO: Huawei Atlas 960E SuperPoD: First NPO/Hi-ONE Optical Node, 4096 NPUs, 8 EFLOPS FP8 · DeepSeek Open-Sources Ascend Versions of TileLang, DeepGEMM and DeepEP as Huawei Details SuperPoD Flex

Beijing-based commercial spacecraft company AZSpace says the structural test model of its DEAR-21 spacecraft has completed mechanical testing. DEAR-21 is the first flight model of the company's C2000 series of medium cargo spacecraft and is planned to launch and fly in the first half of 2027.

Mechanical testing checks whether a spacecraft's structure can survive the vibration, shock and loads of launch. The milestone follows the completion of the structural test model's assembly, which AZSpace announced on October 20. The schedule has moved back from earlier plans: when AZSpace showed a full-scale model of the C2000 demonstration version at a trade fair in Chongqing in May, it said the maiden flight was set for the end of 2026.

The C2000 is a step up in size for the company. Each C2000 spacecraft is designed to carry up to 2,000 kilograms, and several can be combined into what AZSpace calls version 2.0 of a commercial space station. AZSpace has said the C2000 is meant for low-cost delivery of supplies to orbit, reusable orbital transfer, and modular assembly of commercial space station sections. It is designed to stay in orbit for long periods and to support rendezvous and docking.

Founded in 2019, AZSpace focuses on commercial spacecraft and commercial space stations, with stated business lines in space station development, in-space manufacturing and space tourism. Its spacecraft under development include the B300 series of small cargo craft, which carry up to 300 kilograms each and can be combined into an uncrewed commercial space station, version 1.0; the C2000 medium cargo series; and the D6 suborbital crewed spacecraft, designed for six to nine people.

The company's first product, the A10 series, made its first flight in 2023. According to AZSpace, the B300 series completed a full commercial cycle in 2024, and in 2025 the company carried out what it describes as China's first in-orbit operation of a small commercial cargo spacecraft.

Moving from 300-kilogram to 2,000-kilogram cargo craft is the next hard step on AZSpace's path toward a commercial space station. The 2027 flight will be the first real test of whether the C2000 can do the job.

China released its first batch of extreme-weather artificial intelligence models built on Fengyun meteorological satellite data on October 29, at the 2026 Fengyun Satellite User Conference, according to CCTV News. The release is meant to move Fengyun satellite applications from describing the weather as it is toward predicting what comes next.

The first model interprets typhoon cloud imagery. It combines multichannel Fengyun cloud images with related satellite products to automatically analyze a typhoon's position, structure, intensity and likely evolution, then writes up its findings as interpretation text. The model now covers seven categories of typhoon features and draws on more than 20 real-time satellite products. It runs an interpretation every 15 minutes and needs only 27 seconds to generate the text.

The developers used this year's 26th typhoon, Surigae, as an example. After automatically identifying the storm, the model produced a combined analysis of its intensity, wind radius and rainfall characteristics to support monitoring and forecasting. Work of this kind has traditionally depended on forecasters reading satellite images by hand, so automating the first pass can give them a faster, more consistent starting point.

The second model is aimed at earlier warnings of severe convective weather such as thunderstorms. It uses the past two hours of continuous observations from the Fengyun-4 satellite and applies AI to forecast how the clouds will develop. Its key idea is to split cloud evolution into two parts: the overall movement of large weather systems, and the local growth and decay of small ones. The approach extends warning lead time to four hours.

The conference also introduced an artificial intelligence satellite meteorology technology and application innovation team at the Xu Jianmin Meteorological Satellite Innovation Center. The team brings together 49 researchers from 12 institutions and will work jointly on multimodal representation of satellite weather data and on monitoring and early warning of extreme weather.

The satellite models add to a growing set of AI weather tools in China, which so far have centered on medium-range forecasting models run on reanalysis data. These new models work directly from satellite observations, the data forecasters rely on most when a storm is already forming.

SEE ALSO: China Meteorological Administration Puts Fengqing AI Ensemble Forecast Into Operation · China's MAZU AI Weather System to Reach 30 Countries

Alibaba's workplace platform DingTalk has open-sourced DWS, the DingTalk Workspace command-line tool, and released a version that runs on Huawei's HarmonyOS PCs. The project, published on GitHub under the Apache 2.0 license, describes itself as DingTalk Workspace on the command line, built for both people and AI agents.

According to IT Home, DWS spans 25 business modules and more than 1,000 commands. The services listed in the documentation include contacts, chat and group management, calendar events and meeting rooms, to-dos, approvals, attendance records, work reports, AI tables, documents, whiteboards, cloud drive, AI meeting minutes and mail. The idea is that a user describes what they need and an agent carries it out through these commands, within the permissions it has been granted, instead of the user clicking through messages and to-dos one by one.

Agent use is a core design goal. Commands return structured JSON, and the tool ships with built-in agent skills that can be installed per product or as a single combined skill. The documentation recommends a confirmation-skipping flag for agents, a dry-run option to preview a request without executing it, and a filter option to pull out only the fields needed so agents spend fewer tokens.

Because DWS touches company data, it currently requires authorization from an enterprise administrator. The project lists OAuth device-flow login, encrypted token storage tied to the device, a domain allowlist that by default only sends credentials to DingTalk domains, enforced HTTPS, and protections meant to stop agents from being tricked by malicious input, such as path traversal and Unicode spoofing filters. DingTalk says every call passes through its open platform's authentication and audit chain.

The HarmonyOS PC support comes through an OpenHarmony build. The repository produces an openharmony/arm64 binary with an OpenHarmony Go toolchain, and every beta and stable release now builds, self-signs, verifies and publishes a static package alongside the six other platform archives. The project says the binary has been validated on a HarmonyOS PC. Some features are unavailable on that build, including SafeChat and runtime native payloads.

IT Home describes it as the first open-source office command-line tool on HarmonyOS PCs, filling a gap in productivity software on Huawei's desktop platform.

SEE ALSO: Huawei Opens DevEco Studio Public Beta on HarmonyOS PCs · Doubao Work and Tencent WorkBuddy Push Desktop Agents Onto the Phone