|
|
|
|
Bottom Line Up Front Alibaba released an updated Qwen flagship, Qwen3.8-Max-0902, is post-trained for coding and office work resulting in it ranking first on CodeArena's WebDev leaderboard while undercutting leading model's cost at about $5 per million tokens. China's commerce ministry published nonbinding export pricing guidelines yesterday, letting automakers base export prices on cost while asking them to set clear price tiers and add no undisclosed markups. Mifeng delivered its 20,000th capture rig to JD.com Monday, a wearable system that records robot manipulation training data while a person performs ordinary work, and the company says cumulative output has passed 1 million hours. |
| Alibaba's updated Qwen flagship tops a front-end coding leaderboard at about $5 per million tokens Alibaba released an updated snapshot of its flagship model, Qwen3.8-Max-0902, post-trained for coding and for office work. On CodeArena's cost efficiency ranking, the snapshot undercuts every model priced above its average of about $5 per million tokens. On the same site's WebDev leaderboard, which scores the browser-side code that builds what a user sees and clicks, it rose 22 points to 1,691 and ranked first. Among the models it passed are Anthropic's Claude Opus 5 and Moonshot AI's Kimi K3, according to PingWest. Alibaba is serving the snapshot through the programming interface on its Qwen platform. Its Qwen office tools, the Qoder coding agent and the Qwen app all connected to the model at launch. |
| Beijing asks Chinese automakers to set clear export price tiers and drop undisclosed markups China's commerce ministry publicly published policy guidelines yesterday on how Chinese automakers compete in export markets. The instrument allows cost-based pricing and asks for clear price tiers by vehicle configuration, no undisclosed markups on a stated price and respect for host country dealers' pricing autonomy. It also asks companies to keep data handling for connected vehicles and automated driving compliant in the countries they sell in, covering collection, use, protection and cross-border transfer. Separate provisions tell firms to regularize their use of design, component and communications-technology intellectual property abroad and to prepare for disputes. The document describes itself as general guidance for companies' reference rather than a binding rule, and the three agencies share authority to interpret it. |
| Robot-data supplier Mifeng delivers its 20,000th wearable capture rig to JD.com Mifeng Technology's 20,000th MEgo capture rig came off the line Monday, and chief executive Yao Maoqing handed it to JD.com. A capture rig is a body-worn set of cameras and sensors that records how a person's hands do a task, producing training data for robots. Cumulative output has passed 1 million hours, the company says, about half of it bare-hand capture recorded from the wearer's own viewpoint and the rest from wrist-worn rigs and a handheld gripper. The rig pairs panoramic cameras with a wrist camera and aligns its color video, depth, inertial, tactile and audio streams in hardware to under a millisecond. Mifeng also announced a data partnership with Tencent's Robotics X lab, which along with Ant Group's Lingbo already buys from the million-hour set. |
| Horizon's driver-assistance chip reaches more than 70 vehicle models on sale Horizon Robotics' Journey 6M chip was in more than 70 production models on sale as of the end of August, with several automaker brands using it for navigation-guided assisted driving on city streets. A single-chip Journey 6M system supports air cooling, so gasoline cars can meet the heat and production requirements it imposes. Horizon says that extends the capability from a few premium models toward the 100,000-yuan ($15,000) mainstream segment, covering a wider band of prices and powertrains. Rated at 128 trillion operations per second, the part runs a single end-to-end software stack across highway, city and parking within mainstream compute and cost budgets. SAIC Volkswagen's ID.ERA 5S went on sale Aug. 21 as the first vehicle carrying the assisted driving stack from CARIZON, the joint venture between Volkswagen and Horizon. |
|
| Ant Group wins a best paper award for a wide-table system that cut training data prep from 14 days to 2.5 Ant Group's paper on OmniTable won the industrial-track best paper award at VLDB, a leading database research conference, at its Boston meeting yesterday. The system replaces scattered physical tables with one logical table per data domain, so adding a quality feature happens in one place rather than across the more than 100 tables the prior setup required. In production it manages more than 35 petabytes of web, code, PDF and post-training example data, the curated responses used to shape a model's behavior after pretraining. On a real preparation task, Ant reports the end-to-end cycle falling from about 14 days to 2.5 and manual steps dropping by roughly three-quarters. Pre-building copies of frequently read column groups raised filtered-export throughput more than fourfold. |
| Tencent opens its agent platform to outside hardware makers with more than 100 partners Tencent's WorkBuddy open platform went live with more than 100 partners in its first batch. Tencent is selling the layer underneath rather than the app on top. Outside hardware makers, industry software vendors and independent developers can build on the agent foundation the company runs. Nine co-branded hardware products debuted at the launch, from partners including the recorder maker Plaud, the glasses maker Rokid and the camera maker Insta360. Dozens of industry applications connected on day one, spanning finance, law and healthcare. Developers who register and pass a qualification review can build skills, connectors and outside-application integrations on it. Liu Yi, the Tencent Cloud vice president who runs WorkBuddy, said Tencent wants it to become an operating system for the agent era. |
| ByteDance's Doubao assistant gains parallel sub-agents and screen control on Mac ByteDance added parallel sub-agents and Mac screen control to Doubao Work. The Mac feature needs no plugin, programming interface or command line, which lets the assistant drive software it was never wired into. Doubao reads the screen and issues the matching clicks and keystrokes, showing the controlled application live in a floating window with a blue cursor marking each action. Touching the mouse or keyboard pauses the tool and hands control back to the user. The other feature splits a job across sub-agents that run different modules simultaneously, so an organizer agent might divide the screening of 80 resumes among four of them. ByteDance had already shipped the same screen control for Windows, and on Mac a single task can drive several applications at once. |
|
| III | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Tsinghua AI institute's robot model raises its own success rate without updating any weights A model called Zeva raised its cumulative success rate on a manipulation benchmark from 26% in its first round of self-improvement to 73% by the fourth, the researchers report, with no weight updated. That substitutes context for retraining: rather than collect data and fine-tune, the model feeds a record of what its own attempts caused back into an action generator whose weights never change. Zeva comes from Tsinghua University's Institute for AI Industry Research and from Domain Transform, a company two of its professors founded. Its causal transition encoder learns which actions produce which changes, and a two-timescale memory pairs a trace within one attempt with a persistent record carried across attempts. In a real chemistry lab the team ran the same test, with cumulative success on picking up a test tube rising to 100% across repeated attempts. |
| Sensor maker Foison delivers more than 30,000 sets of fingertip touch sensors to Linkerbot Foison New Material said Monday it had cumulatively delivered more than 30,000 sets of tactile sensors to Linkerbot, which builds multi-fingered robot hands. That is about a third of the 100,000-set framework the two signed in February. The sensors go mainly into the Linker Hand O6, Linkerbot's industrial-grade dexterous hand. Foison pairs high-precision sensing units with a microprocessor so that each fingertip carries its own signal-processing unit. It has also co-developed a tactile data collection glove with Haocun Technology, distributing several hundred sensing points across the fingertips, finger pads and palm for whole-hand force sensing. |
| Robot maker AI2 Robotics puts a cocktail-mixing robot behind a Hong Kong bar AI2 Robotics said Monday that its AlphaBot 2 had begun mixing cocktails and talking to customers at a bar in Lan Kwai Fong, Hong Kong's nightlife district. AI2 Robotics says it handled export and customs clearance itself, obtaining a local beverage sale qualification and set up a paid service arrangement, laying the farmwork for future commercial deployment. AlphaBot 2 runs NeuroVLA, a model that splits control between a module for task planning, one for high-frequency motion correction and one for fast safety response. AI2 Robotics reports its own experimental data showing the design cuts motion jitter by more than 75% and completes a reflex response within 20 milliseconds of a collision. Chief executive Guo Yandong was pictured alongside Lan Kwai Fong Group chairman Allan Zeman at the announcement. |
|
| IV | QUANTUM & FRONTIER SCIENCE | |
| State automaker FAW cycles a solid-state battery cell at a quarter of the usual working pressure China FAW said it had cycled a large-capacity all-solid-state battery cell stably at 5 megapascals of working pressure, working with the Eastern Institute of Technology. The obstacle is mechanical rather than chemical: with a solid electrolyte in place of a liquid one, all-solid-state cells generally need more than 20 megapascals of working pressure to keep electrode and electrolyte in contact. FAW's figure is a quarter of that, and the company says the reduced dependence on external pressure suits pack-level engineering design. FAW attributes the result to controlling the morphology of the materials, a low-pressure structural design for the cell and an optimized fabrication process. The work appears in Advanced Energy Materials journal." |
|
| Robot maker Astribot hires ByteDance reinforcement-learning lead Sun Peng to head robot training Sun Peng joined the robot maker Astribot to lead its robot reinforcement learning research and deployment, coming from ByteDance. The hire moves the problem from whether a robot can do a task to whether it does so reliably, which is a reinforcement learning question rather than a pretraining one. Reinforcement learning trains a system by rewarding the outcomes of its own attempts rather than by copying labeled examples. At ByteDance he led ByteRL, the company's core reinforcement learning infrastructure, and before that ran the agent center at Tencent's Robotics X lab. Astribot builds its stack as foundation models, an embodied operating system and a cable-driven robot body, with its Lumo-2 model predicting future states of the world before generating an action. |
|
| Chinese outlet translates a June CNAS report urging 72-hour reviews of Chinese AI releases A Chinese translation of "Red Lines," a June 12 report by the Center for a New American Security, circulated on Chinese tech sites today. The translation carries an editor's note calling the report U.S. scenario-building rather than objective fact. The report asks the U.S. Commerce Department to publish a national security risk assessment no more than 72 hours after a major Chinese AI system is released. On tests of AI agents solving software engineering problems the US still leads with Chinese models trailing the top U.S. score by under 10 points. However, Chinese models lead in diffusion, with the three most-used models on OpenRouter all Chinese as of early 2026 and priced up to 17 times cheaper than comparable U.S. models. Chinese systems are not yet deployable in weapons platforms, the report says, because onboard power draw is too high and real-time control conflicts with network-dependent inference latency. |
|
|
|