Huawei's HarmonyOS Interconnect app is now available for Apple Watch, letting the smartwatch pair with Huawei phones to relay messages and help locate devices. The watch app requires watchOS 10.0 or later, and the Huawei phone must run HarmonyOS 7.0.0.109 or later.

For now, only three Huawei phone lines can connect: the Pura X View, the Mate 90 series and the Mate XT 2. Huawei said on its official support site that other models will gain the feature gradually.

Setup runs through the watch. Users download HarmonyOS Interconnect from the App Store on Apple Watch, make sure both devices have Bluetooth on, are online, unlocked and close together, then open the app on the watch to display a QR code. Scanning that code with a Huawei phone signed in to a Huawei ID completes pairing. Once connected, notifications can flow between the devices, and each can be used to find the other.

On the App Store, the app is listed under developer Thunder Software Technology and also runs on iPhone, iPad and Mac. On those devices it supports lossless file transfer with Huawei devices on HarmonyOS 6.0.0.112 or later without using mobile data, and files from devices signed in to the same Huawei ID can be received without confirmation. Version 2.0.0, released after Huawei's October 1 launch event for the Mate 90 series and other products, added a dual-device interconnection mode that pairs an Apple device with a Huawei phone for message relay, device finding and network sharing. That feature likewise launched first on the Pura X View, Mate 90 series and Mate XT 2.

Huawei first said the app would support Apple Watch and AirPods at its September 7 launch event for the Mate XT 2. For AirPods, Huawei has described showing earbud battery levels on the leftmost home screen panel of a Huawei phone.

The Apple Watch release is the first concrete step in that cross-ecosystem plan. It targets a practical gap for people who own a Huawei phone and an Apple Watch, a pairing that Apple's own software does not support. Huawei has not given a timeline for adding more phone models.

SEE ALSO: Huawei Mate 90 Pools Networks of Up to Four Phones, Cutting Video-Call Stutter 90% in Rail Test · Xiaomi HyperOS 4 Opens Up to Apple Devices With AirDrop Transfers, Notification Sync and Mac Control

Several of China's biggest internet companies are building AI tools for interactive film games, a genre of branching, playable video stories, in a bid to own the creator entry point for what could become a mass-produced content format, according to a Chinese media report published on October 11.

On September 20, ByteDance's video editor CapCut, in its China edition, unveiled ICG Studio, an AI interactive film game creation platform now in closed beta. Creators start from a story outline, then generate a script, persistent character, scene and prop assets, and storyboards. By chatting with an agent, they can branch the plot at any decision point and generate new storyboard shots, so characters and sets can be reused across storylines.

Tencent has four products across different teams. TDream, billed as AI engines for playable cinematic worlds, turns scripts into video and adds branching choices. LightSpeed Studios launched a creation platform in May that breaks interactive story production into six standardized steps for non-professional users. Project Craft, an AI game creation platform shown at Tencent's SPARK 2026 event, generates 2D and 3D game prototypes from text. DreamNow is a showcase and discovery platform for AI videos, images and interactive titles. Kuaishou said in July that AI interactive content creation features were coming.

Hits such as The Invisible Guardian and Love Is All Around showed there is an audience for the genre, but high production costs have kept the supply of quality titles thin. Developers interviewed say AI is already cutting costs. One game producer said character art that used to cost RMB 2,000 per image now costs about RMB 800 with AI assistance, and a RMB 200,000 small-game budget can fall to around RMB 150,000.

Problems remain. Interactive scripts are far longer than short-drama scripts, and AI still cannot write them well. A hoped-for data flywheel, in which every player choice trains better plots, is unproven because the reasons behind a click are hard to attribute. The genre also lacks a dedicated distribution channel; most titles still sell as standalone games on platforms such as Steam, and whether AI-made titles will need a game publishing license is unclear.

For now, the near-term opportunity looks strongest in tools for professional creators, which helps explain why the platforms are racing to offer them.

SEE ALSO: Tencent Open-Sources Hunyuan 3D World Model 2.0 for Interactive Content Creation · Kuaikan Manhua Expands into AI Interactive Content, Targets May Beta Launch

A developer in the DeepSeek Harness community has released DS Bot, a plugin that turns DeepSeek's open-source agent runtime into a team of persistent bots that users can talk to one by one or in group chats. The project is not an official DeepSeek product. It was published on GitHub by a developer who goes by FeiZ and picked up more than 250 stars within a day of launch, according to QbitAI; the repository had passed 450 stars by October 11.

DS Bot follows the everything-is-a-plugin design of DeepSeek Harness, or DSH. Users create bots and give each one a name, a role description, instructions, an avatar and its own underlying model. Any model already configured in DSH can be used, and the author says local agents may be supported later. Models can be switched at any time, although the author recommends sticking with one model per bot for the long run.

The plugin's answer to rising token bills is a main bot. On first launch a default main bot called Chief is already in place. It has elevated permissions to create and configure other bots and group chats, and the author recommends doing setup simply by talking to it. Every group chat also has an admin bot that replies to the user by default and calls in other bots only when needed, instead of having every bot answer every message. Users can still mention a specific bot to talk to it directly.

Memory is the other focus. Each bot maps to its own DSH session with its own working context. When a conversation outgrows the context window, DS Bot falls back on compression, handover summaries and local search. Each bot keeps Markdown-based memory files, there is a shared team memory, and after every turn the model reviews what should be saved, updated or deleted. Memory can be edited on a bot's details page or by telling the bot.

A usage page breaks down token consumption by time, bot and model. All data stays on the user's computer in the bot and sessions folders of the DSH home directory, and uninstalling the plugin does not delete it. The project is licensed under Apache 2.0 and is labeled a v0.1 preview built against DSH 0.2.1-alpha.1.

DS Bot installs from the Plugins panel in the DSH desktop app on Windows and macOS, or through the DSH web profile, which needs Node.js 22.19 or later and pnpm. It joins a growing list of community multi-agent plugins for DSH.

SEE ALSO: DeepSeek Harness v0.2.1-alpha.2 Pre-Release Adds Reasoning Translation and Git Worktrees · DeepSeek Harness v0.2 Preview Adds Desktop Installers, Plugin Manager and Scheduled Automations

Ant Group has quietly started a limited beta of a new version of LingGuang, its multimodal AI assistant, and is repositioning the app as a personal agent along the lines of Meta's Muse, according to Sina Tech, citing people close to the project. Ant has not made an official announcement.

Meta introduced Muse in September as a personal agent that steps into users' daily lives and keeps following up on the things they hand over. According to the sources, Ant began preparing a similar direction for LingGuang about six months ago, iterated through several internal versions and plans to open the new release to all users this month.

The core change, the sources said, is a shift in what LingGuang is for. The current app is built around chatting and generating content and small applications. The new version is designed as an AI-native personal agent with long-term memory that can move tasks forward on its own initiative rather than waiting for each prompt. It can also call Skills and Model Context Protocol (MCP) tools, which lets it send messages through Alibaba's workplace app DingTalk and operate smart home devices.

The beta use cases focus on long-running jobs with many moving parts that ordinary users struggle to keep track of. Examples cited include monitoring market swings for office workers, helping parents manage check-ins on DingTalk, and reworking a travel route when a flight is delayed or a scenic site starts limiting visitor numbers.

LingGuang launched in November 2025 as what Ant described as China's first multimodal AI assistant to answer through code-driven outputs. It can respond with 3D models, audio clips, charts, animations and interactive maps, and its Flash App feature builds small working applications from a natural-language prompt in as little as 30 seconds. Users can publish those apps to an in-app community called LingGuang Circle. A web version later extended the same features to desktop browsers.

A move from on-demand generation toward persistent, proactive agents would put LingGuang in a crowded field. Several large Chinese internet companies are now working on personal assistants that remember context across weeks and act across apps, and the main questions for each are reliability on long tasks and how much access users are willing to grant. Details such as pricing, data handling and which third-party services the new LingGuang will connect to at launch have not been disclosed.

SEE ALSO: Meta and Tencent's Two Paths to the Personal AI Agent · ByteDance's Doubao Reportedly Readies Personal AI Assistant From Project Codenamed Spell

A chip startup founded in 2025 by Xu Feixiang, a former GPU architect at Nvidia, is designing a processor built from scratch for physical AI, the robots and machines that sense, reason and act in the real world. Its first product, Y10, aims to cut the full perception, inference and action loop to under one millisecond, Xu told Chinese startup outlet Cyzone in an interview published October 10.

Xu argues that most robots today make do with CPUs, GPUs and autonomous-driving chips. GPUs are built for throughput on large language models, he said, not for the continuous loop in which a robot senses its surroundings, acts and then adjusts at high frequency based on feedback. For physical AI, the useful measure is how long the whole loop takes and what it costs in power and money, not just peak TOPS. He estimates that chip architecture accounts for about 70% of how usable a parallel processor is, and software for 30%.

Rather than a fixed-function ASIC, Y10 is a general-purpose processor driven by a custom, highly programmable instruction set, so it can keep up as robot models change. According to earlier company material, the design is based on an architecture it calls SomaArch. Xu said core team members worked on several generations of Nvidia GPUs and took large chips, including 6-nanometer parts, from design to mass production, and that the instruction set is designed to ease migration for engineers.

The company says it is testing with leading Chinese robot makers. Against the x86 plus Orin setups customers use today, its simulation results show five times the overall performance at 80% lower cost. These are the company's own simulated figures, not measurements on silicon; Xu said the team's simulations usually land within 3% of final results. Chip development is past the halfway mark, and an FPGA version of the computing platform is planned for the end of this year so customers can test real hardware.

The first target is industrial robot arms rather than humanoids. Citing a partner, Xu said existing arms handle only about 2% of manual operations on real production lines. The company is also considering small world model servers built on its chips. Today's robot computing power, Xu said, may be roughly where the 386 processor stood in 1985.

SEE ALSO: Infinigence AI Open-Sources APXInf for Embodied Edge Inference on Jetson Thor · Moore Threads MTT S5000 Embodied RL Curves Match Mainstream GPUs (r=0.976); Dual-Arm ~92%

Westlake Robotics, the Hangzhou robotics company spun out of Westlake University, has released WR1, which it calls a general brain for humanoid robots. In demonstration footage published on October 9, a humanoid running WR1 takes corn from a fridge, makes a smooth U-turn to an air fryer, squats to drag a storage basket and tidy scattered plush toys, then walks to the bedroom to straighten a quilt, without stopping to reset between rooms.

The company says WR1 rests on a dual pretraining design for a large brain and a small brain. The brain combines a world model with a vision-language-action model to perceive the scene, understand the task and output one whole-body action sequence covering walking, torso posture and dexterous hand movements. A general action expert called GAE acts as the motion cerebellum, tracking that sequence and keeping the robot balanced. GAE was trained in two stages on large volumes of human motion data and uses an adjustable look-ahead to compensate for hardware delay.

To stop long tasks from breaking apart, WR1 keeps a short- and long-term memory of the robot's motion and generates actions in chunks. The next chunk is computed in the background while the current one is still running, so updates do not cause jerks. Westlake Robotics argues that most humanoids plan locomotion, posture and hand work separately and must stand still before they can use their hands.

The company's main claim is about laundry. In the demo, a humanoid lifts a pair of trousers from a swaying clothes rack, drapes them over its arm while adjusting to the movement, and works with a second WR1 humanoid to pass, spread and fold them with dexterous hands as both robots track each other and the changing fabric. Westlake Robotics calls this the first public demonstration of a humanoid fully folding clothing with dexterous hands, a claim that has not been independently verified. Earlier, WR1 folded clothes on a robotic arm platform, a separate setup the company says reached 36 seconds per item at its fastest.

Westlake Robotics was founded by Donglin Wang, a tenured professor at Westlake University, with co-founder Yue Zhang, a natural language processing researcher and tenured professor at the same university. It develops its brain, whole-body control and humanoid hardware in-house and targets home service, manufacturing and specialized work.

SEE ALSO: Light Origins Open-Sources Light-O1-Preview 6B Whole-Body Model · Robotera's VPP2 World Action Model Tops RoboDojo, Scores 58.5% Zero-Shot on Real ALOHA Arms

DeepSeek has published DeepSeek Harness v0.2.1-alpha.2, a new pre-release of its open-source agent runtime. The build went up on the project's official GitHub page on the evening of October 9 Beijing time, six days after v0.2.1-alpha.1, and is still marked Pre-release rather than a stable version.

The headline addition is an experimental plugin that machine-translates a model's reasoning output. Users switch it on from the Plugins page and choose Bing, Google or paid DeepSeek Flash as the translation provider. A second experimental plugin adds Git Worktrees, letting an agent create and move into a separate Git checkout so parallel work does not collide in one working copy. Experimental plugins also get new session state-record APIs for writing and reading their own records.

Several changes target developers building on the runtime. A new working_directory tool lets the agent read and change the session's working directory, and the TypeScript and Python SDKs gain matching APIs. Global instructions can now be loaded from an AGENTS.md file in a shared agents directory, set through the DSH_AGENTS_HOME variable. The Official plugin catalog adds on-demand installation of Claude Code and Codex bundles that match the running release, and these bundles can also be used in headless, SDK and ACP profiles. SSH gains standalone helper runtimes that handle remote file operations, processes, terminals, sandboxing and Node PTC.

The web version can now bind to a specific IPv4 or IPv6 address and serve HTTPS directly through new --tls-cert and --tls-key options. Other additions include font and size settings, microphone selection for voice input, and a choice of when to collapse work details.

Among the fixes, requests no longer hang past the idle timeout when a model stream stops and the transport ignores cancellation, and background workflows are no longer marked interrupted while still running. A failure to read plugin metadata no longer blocks an entire DeepSeek request.

The release also carries breaking changes. The mixed both tool-presentation mode is removed, leaving native and ptc. Subagent tools now return a child ID immediately, so custom SDK runtimes must support session/wait. Agent Team messages go straight to the target inbox, dropping the separate outbox and automatic retries. The default SDK profile changes from a fixed coding-agent persona to a general-purpose AI agent identity, and the bundled pi-ai library moves to version 1.0.2.

SEE ALSO: DeepSeek Harness v0.2.1-alpha.1 Adds Experimental Claude Code Mods Compatibility Layer · DeepSeek Harness v0.2 Preview Adds Desktop Installers, Plugin Manager and Scheduled Automations

Maniformer, a Shanghai company that supplies training data for physical AI, has opened a crowdsourcing platform that lets ordinary people earn money by recording everyday tasks for robots to learn from. The company launched the platform at an event in Shanghai on September 23 under the slogan that anyone can be a teacher, Leiphone reported. It calls it the first all-category crowdsourcing platform for high-quality physical AI data.

Maniformer CEO Yao Maoqing argued that what keeps robots out of factories, shops and homes is not algorithms or hardware but a lack of large volumes of real human motion data. The platform aims to turn routine actions such as folding clothes, making tea or arranging goods into usable training data.

It works in two layers. The base layer is infrastructure: MEgo capture devices, a mobile app and the MEgo Engine for post-processing. MEgo View is a head-worn rig whose cameras cover more than 300 degrees, and MEgo Gripper reconstructs motion to millimeter accuracy with optional 3D touch sensing. On top sits a network of real work sites and data collectors. After registering under their real names and receiving a device, users take a job, record it, pass a quality check and get paid, with withdrawals said to arrive within two minutes and taxes handled by the platform.

The platform lists 22 categories, more than 5,000 task types and over 50,000 real scenes. During a closed test from August 15 to September 15, nearly 20,000 users signed up.

Maniformer also launched a robot trainer program and gave the title to 16 beta users. Its business vice president Zhang Zhifu outlined a RMB 100 million subsidy plan: RMB 50 million for tasks, RMB 20 million for devices, RMB 30 million for an offline service network in 40 cities and RMB 2 million for insurance covering people and equipment. A five-level M1 to M5 progression system comes with a goal of 1 million daily active users and 1 million robot trainers within two years.

The company is also working with the China Academy of Information and Communications Technology on a report on crowdsourced embodied AI data, has set up a scene data alliance with 12 initial members, and issued a data security initiative covering collection rules and end-to-end protection. It says the platform already operates in 40 cities.

SEE ALSO: Maniformer Launches One-Stop Physical AI Data Platform to Power the AGI Era · X Square Robot's TwinDEX Learns a 24-Step Chemistry Experiment With Zero On-Robot Training Data

Striding AI, a Beijing physical AI startup, has unveiled a robot system for convenience stores that run around the clock. The company showed it on October 8 at the 22nd Asia-Pacific Retailers Convention and Exhibition (APRCE 2026), according to a company release carried by QbitAI.

The system has three parts. H1 is a bipedal humanoid designed for spaces where customers move around. It speaks more than 20 languages, can show 45 facial expressions and 36 gestures, and offers an open API for custom development. C1 is a wheeled robot with arms for narrow aisles. It runs on a slim omnidirectional base, reaches from about 100 mm to more than 1,750 mm high, and handles moving goods, restocking shelves and loading. Striding AI says C1 senses nearby shoppers with fused sensors, stops within 10 cm of an obstacle in an emergency, and carries front and rear depth cameras plus 360-degree vision. M1 is the management platform that assigns tasks, tracks status and lets staff take over.

The company says the setup needs no changes to store shelves or walkways, which it calls zero retrofit. It wants the robots to take on repetitive and overnight work such as restocking, stock checks and patrols, so staff can focus on customers. It claims robots complete 99% of tasks on their own, a figure it has not explained in detail.

Behind the robots is what Striding AI calls its embodied brain. SLM-0.5, a lightweight world action model, scored a 98.6% average task success rate on the LIBERO benchmark and runs inference 24 times faster than mainstream approaches, the company says. STEAM, a reinforcement learning post-training system built on the open-source RLinf framework, reached 95% success on a specified shelf-stocking task. Rpent, an agent layer launched with Tsinghua University and Infinigence AI, breaks a request such as filling a row of shelves into steps, calls navigation and grasping modules, and replans when something slips.

Striding AI was co-founded by serial entrepreneur Song Yao, Yang Yuxin, Tsinghua researcher Yu Chao and the Charoen Pokphand Group. It plans to test the system with a large global retail chain and to start commercial service in 2027, sold outright or as a Robotics-as-a-Service subscription covering hardware, software and maintenance. A heavy-duty version and an A1 robot for checkout and bagging are planned, without dates.

SEE ALSO: Infinigence AI Open-Sources APXInf for Embodied Edge Inference on Jetson Thor · Moore Threads MTT S5000 Embodied RL Curves Match Mainstream GPUs (r=0.976); Dual-Arm ~92%

Alibaba's AI coding agent Qoder has added a Fast mode built for everyday tasks that need quick turnaround. The company said on October 10 that the new tier is live first in the Qoder desktop app, version 0.4.4 and later, for both individual and enterprise users, on the China and international editions at the same time, IT Home reported.

Qoder lists Fast with a first response time of about three seconds, high output quality and a credit rate of 0.8x. On the China edition, Fast is a new tier added alongside the existing options. On the international edition, it replaces the previous Performance tier rather than sitting next to it.

For international users, the company says the change brings a faster and cheaper tier with no loss in quality. First response time falls from about eight seconds to three, and the credit rate drops from 1.1x to 0.8x, roughly 27% less. Qoder says output quality is the same as Performance, and users on the old tier are moved over automatically with nothing to change on their side.

Qoder also published first-response times for its other tiers. By its own measurements, Fast at three seconds compares with 7.9 seconds for Performance on the international edition, 9.8 seconds for Auto and 13 seconds for Ultimate on the international edition. That puts Fast at under a third of Auto and roughly a quarter of Ultimate.

The company also ran its own comparison against the fast modes of rival coding tools. It said it picked ten of the most common developer tasks and timed how long each product took to finish answering. In that self-reported test, Qoder said it finished in about a third of the time of comparable domestic products and about a quarter of the time of comparable overseas products. It did not name the products tested, and the figures have not been independently checked.

Example tasks shown in the announcement were light: drafting a short travel itinerary in about a second and turning a daily work log into a plan for the next day in about four seconds.

Qoder said other products in the Qoder family will add Fast mode in the near future, without giving dates. The tier is aimed at quick, routine work, while slower tiers such as Ultimate remain available for heavier jobs.

SEE ALSO: Alibaba Launches a Rebuilt Qoder Coding Agent Aimed at Non-Developers · MiniMax Puts M3.1-Flash-Preview Coding Model Live Inside MiniMax Code

The AI research team MirroS has released AgentGarten, a framework that gives AI agents a world they can act in, observe and learn from. QbitAI reported on October 9 that the code, a technical report and a project page are all public.

AgentGarten splits the job of simulating a world in two. Code handles physics, layouts, contact and game rules, so the state is explicit and every outcome is decided by the program. A neural renderer then turns a light geometry sketch exported by the code, such as depth or surface normals from the agent's camera, into a realistic next frame. The team argues this avoids two common trade-offs: game-engine scenes that need costly art work to look real, and video world models whose physics is hidden inside the network and cannot be queried or edited.

The renderer generates video in short streaming chunks, so an agent can act, see the result and decide again. MirroS trained it with a method it calls Adversarial Forcing, which adds exact replay so gradients reach the model's generated history, and a discriminator trained on real video to curb drift and grid-like artifacts over long rollouts. Custom Triton kernels, CUDA graphs and a lightweight decoder keep it above 30 frames per second at 480p on a single GPU.

To test learning, the team revisited the hide-and-seek game from a well-known 2019 study, in which reinforcement learning needed about 25 million episodes before hiders built shelters and about 100 million before seekers used ramps. In AgentGarten, agents saw only rendered first-person frames and acted through short Python programs. After each round they reviewed their games and wrote lessons into a playbook passed to the next round. Hiders built shelters by round four and seekers climbed over walls with a ramp by round ten.

The same loop ran unchanged in four other worlds over four rounds each. A companion-dog score rose from 13 to 19, two cars crossing a one-lane bridge cut their total time from 71 to 41 seconds, two herding dogs went from three sheep penned to all four, and a quarry loader finished its job 31 seconds early.

MirroS frames the work as a step toward agents that keep improving in physical settings. It is a research environment, not a deployed robot product.

SEE ALSO: AIsphere Launches PixVerse R2, a Real-Time World Model That Remembers What Happens in a Session · ACE Robotics and NTU Open-Source Puffin-World Multimodal World Model

ShengShu Technology has released Vidu Q4 Preview, the first preview of its next flagship video generation model, with two selling points: more convincing character performances and a much lower entry price. IT Home reported the release on October 10, and a hands-on review appeared on Tencent News on October 8.

The model accepts up to 15 reference images, giving characters, scenes and props a fuller visual reference, and up to three reference audio clips to keep a character's voice consistent. ShengShu says it improves dynamic camera movement, shot changes and transitions between complex shots. Output runs at 540p, 720p, 1080p, 2K and 4K with 10-bit color depth, and a single generation can last up to 16 seconds. For now, Q4 Preview supports image-to-video and reference-to-video generation only.

Pricing is the other headline. For Vidu's own SaaS product, the preview carries a two-month launch offer starting at RMB 0.09 per second, which puts a 10-second clip below RMB 1. Neither report gave the standard price that will apply after the offer ends. The reviewer noted that mainstream video model APIs in China typically charge between RMB 0.5 and RMB 3 per second, and that creators often generate many takes before getting a usable clip, so per-second cost largely decides who can afford to experiment.

In the reviewer's tests, which included a fantasy advertisement, an anime short and a film-style scene, the model handled a slow pull from a close-up of a dessert plate to a wide garden landscape without the abrupt zoom-style cuts common in earlier models, and kept cuts between shots continuous. The reviewer described a recognizable house style favoring energetic camera moves, anime and cinematic action. These are one reviewer's impressions, not benchmark results.

ShengShu pitches Q4 Preview at AI short dramas, advertising, social content and film production. The Q3 generation was aimed at getting AI characters to act consistently in production workflows. With Q4, the company is pairing better performances with a price low enough to encourage repeated drafts. The full Q4 release date has not been announced.

SEE ALSO: Vidu Claw Launches "100 RMB for Million-Dollar Commercials" Era: Prompt a Video in WeChat, Get a Theater-Grade Ad · Shengshu Technology Launches Motubrain World-Action Model

Robotera, the Tsinghua University-backed robotics company, says its VPP2 world action model now ranks first on RoboDojo, a simulation benchmark led by the University of Hong Kong's MMLab with nearly 20 academic institutions. According to QbitAI, VPP2 posted an average success rate of 32.26% and an average score of 39.26 across RoboDojo's 42 dual-arm tasks, which test generalization, precise manipulation, long-horizon tasks, memory and open-vocabulary instructions. It also ranked first in the generalization, precision and memory categories.

Robotera says VPP2 reached those numbers without extra training data or add-ons such as agent-based recursive self-improvement, using only the standard dataset. The strongest baseline cited in its paper, GPT-6-Astra in post-training evaluation, scored 22.48% and 28.97. The ranking reflects the leaderboard figures cited in the paper. Simate made a separate claim to the top of RoboDojo for its Simate-beta system last month.

The model targets a known weak spot of world action models: if the video prediction is wrong, the action follows it. Robotera splits training into three stages. First, event-level video pre-training, built on Alibaba's open-source Wan2.1-I2V-14B, teaches the model to predict an entire manipulation from start to finish, using finely captioned clips from robots, humans and general video. Second, prediction is trimmed to fixed eight-second clips and distilled into a single step, taking about 0.12 seconds. Third, a 0.9-billion-parameter Action DiT learns to turn predictions into motion while the video model stays frozen and adapts only through LoRA. Total latency for an action chunk is about 0.22 seconds.

On a real ALOHA dual-arm robot, VPP2 averaged 58.5% across 10 zero-shot task types, including picking, placing, stacking, folding and pouring, against 40% for Physical Intelligence's pi0.5, and was best on nine of them. It scored 45.0% on LIBERO-Pro and 63.9% on LIBERO-OOD. Adding a vision-language model as a high-level planner lifted selected RoboDojo long-horizon tasks from 27.6% to 57.6%.

Robotera already runs robots in more than 10 logistics centers across five provinces and cities with partners including China Post and SF Express, though that does not mean VPP2 is deployed at scale. The code is open source on GitHub.

SEE ALSO: Simate Debuts Simate-beta Physical AI Fast System, Says It Tops RoboDojo Benchmark · Xiaomi Open-Sources Robotics-U0 Embodied World Model and Training Stack

After a National Day holiday dominated by talk of Huawei's new flagship phones, the more interesting contest in the country's phone market may be over who owns the AI layer that sits between users and their apps. A commentary from the tech column Internet Jianghu, published on Tencent News on October 9, sums up the four biggest players in one line: Huawei goes deep, Alibaba goes broad, ByteDance goes focused and Tencent goes clever.

Huawei is the only one building the full stack itself. Its own chips, its own operating system and its own models let it push AI into the lowest layers of the phone. The column points to the stronger NPU in the Kirin 9050 Pro, which can run a 30-billion-parameter mixture-of-experts omni-modal model on the device. Huawei relies on its own models and connects to WeChat's assistant only through the A2A agent protocol. The upside is a higher ceiling, including offline use and lower power draw. The cost is a very long supply chain that few rivals could copy.

ByteDance has taken a narrower route with the Doubao phone, built with ZTE's Nubia on the second-generation NaviX Ultra. The new version calls apps through the MCP and A2A protocols instead of simulating taps on the screen, but fully automated actions still mostly reach ByteDance's own services.

Alibaba is selling a platform. Its Qwen Intelligence solution gives phone makers model capability while they contribute the system layer and real users, as Honor has done on its Magic9 series and Robot Phone. If the model works there, it could be repeated across other Android brands at relatively low cost.

Tencent is not competing for the system-level agent at all. Through WeChat's A2A partnerships with Huawei, Xiaomi, Honor, OPPO and vivo, phone assistants send instructions to WeChat, which carries them out. Separately, it is working with Qualcomm to run 7B and 3B Hunyuan models on Snapdragon 8 Elite devices.

The column argues the real barrier is the ecosystem. Phone makers fear becoming pure hardware pipes. Model companies gain API revenue but not the user relationship. Super apps such as WeChat, Taobao and Meituan are both the services agents must call and potential rivals, and the first Doubao phone ran into blocks from major apps. The author's conclusion is that the industry needs a new way to split revenue, based on completed user intents rather than ad traffic or token counts, before any company can claim the phone's first AI entry point.

SEE ALSO: Alibaba Launches Qwen Intelligence Phone-Agent Stack; Honor Magic9 Named First Formal Adopter · ByteDance Doubao Phone Agent Goes MCP/A2A-First With GUI Fallback

A coding agent built by Lenovo's Tianxi AI team has taken first place on the Lite leaderboard of SWE-bench-Live, a benchmark that asks AI systems to fix real issues pulled from GitHub projects. Running DeepSeek-V4.1-Flash as its underlying model, Lenovo's TianxiCode framework resolved 71% of the tasks and passed the benchmark's official verification, according to a Lenovo statement carried by QbitAI on October 9.

The result applies to the Lite split only, not the benchmark's full or multi-language boards. The public leaderboard lists the entry, dated October 8, as resolving 213 of 300 Lite tasks. The margin is narrow: the next verified entry on the same board sits at 70.33%.

SWE-bench-Live differs from classic coding tests that check syntax or single functions. Each task comes from a real GitHub issue, and a system must produce a patch that fixes it inside a reproducible execution environment. To earn the Verified mark, teams submit the full trajectories of their agent runs, and the organizers check the evaluation inputs and isolation setup for any leak of reference answers, test cases or results.

Lenovo describes the split of work this way: DeepSeek-V4.1-Flash acts as the engineer's brain, reading code and proposing fixes, while TianxiCode supplies the eyes, hands, tools and working discipline. The company highlights three capabilities. Multi-hop retrieval across files, combined with context pruning and slicing, helps the agent trace a bug to its root cause in a large codebase. Autonomous planning with multi-turn tool calls lets it draft a debugging plan, run tests in a terminal, read logs and compare change histories, then switch approach when stuck. Closed-loop patching with test-driven self-correction runs new code in an isolated environment and keeps revising until the patch passes.

TianxiCode is the code-generation and software-engineering component of Lenovo's Tianxi AI ecosystem. Lenovo says it plans to fold the work into developer toolchains and into its AI hardware products, but it has not given a timeline or said which devices will carry it.

The result also shows how far a cost-efficient Flash-class model can go when paired with a strong agent harness. Lenovo did not publish separate scores for the same framework on other models.

SEE ALSO: DeepSeek-V4.1-Flash Ships Causal Encoder–Decoder MoE With 1M Context and Extreme KV Compression · Lenovo Unveils Tianxi AI 4.0 Ecosystem: Palm-Sized AI Host, 95% Token Cost Reduction, First Large Foldable Phone

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