|
|
|
|
Bottom Line Up Front Tencent released open weights for Hy4 preview, an early version of its large language model that holds 770 billion parameters but activates 49 billion of them per token. More than 20 operators joined China's compute scheduling platform, and the pilot system now monitors 1.45 million PFLOPS of AI compute, routing jobs to hubs with spare capacity rather than to the buyer's own province. Sugon released ParaCache, a shared cache pool that lets one server reuse model state another produced, which cut time to first token by more than 80% on the company's own multi-turn test data. |
| Tencent releases open weights for a 770-billion-parameter AI model Tencent published open weights for Hy4 preview, an early release of its Hy4 large language model. The model holds 770 billion parameters but activates 49 billion of them on any given token, so the model runs more efficiently than one that activates the full parameters. Open weights mean the files can be downloaded and run by anyone rather than reached only through a paid interface, and Tencent has posted them on Hugging Face, GitHub and ModelScope. Its context window runs to a million tokens. In a blind comparison graded by Tencent's own experts across about 200 engineering tasks, Hy4 preview finished less than a tenth of a point ahead of GLM-5.3 and Kimi K3 on a four-point scale. Tencent calls the release an early iteration and names overlong deliberation and excessive self-verification on complex tasks as known problems. |
| More than 20 operators join China's national compute scheduling platform More than 20 compute scheduling operators have joined the pilot platform that monitors and schedules China's national computing power network, the National Data Administration said. Scheduling here means routing a job to whichever hub has spare capacity rather than to the province the buyer sits in, which turns separately owned accelerators into one usable pool. The platform now monitors 1.45 million PFLOPS of AI compute, a PFLOPS being a quadrillion calculations a second, measured at the reduced numeric precision most training and inference run at. The administration sets national data policy and runs the network with the National Development and Reform Commission, China's top economic planner. The 15th Five-Year Plan names the network a major project. Both agencies say they will build it on eight national computing hubs, sited to put data processing near cheaper power in the west. |
| Sugon pools cached model state across servers to relieve an inference memory bottleneck Sugon released ParaCache, a shared cache pool for large language model inference that spans GPU memory, server DRAM, solid-state drives and distributed storage. Sugon's argument is that inference has run into a wall of memory rather than of accelerators, which moves the fix from buying more GPUs toward coordinating storage and compute. The state in question is the KV cache, what a model has already computed about the text it has processed, and without it the model redoes that work for every new token. ParaCache lets one GPU or one server reuse cache that another produced, on the company's test data raising token throughput as much as 27 times under heavy concurrent load. Time to first token, the delay before a reply begins, fell by more than 80% in multi-turn sessions. Sugon validated the system on Shuguang 8000, which it calls China's first fully domestic 100,000-card AI supercluster, completed in July. |
| X Square Robot's world model predicts which robot policies will work on hardware X Square Robot released WALL-SS yesterday, a world model for testing robot control software before it runs on a machine. A world model predicts how a scene changes in response to an action, so if it ranks control policies correctly, a developer discards the weak ones without touching hardware. Running the same policies in simulation and on real robots across hundreds of matched trials, the company found success rates correlated at 0.926, close to a one-to-one match. The model ranked policies by strength correctly 89% of the time. GeekPark sets the release against Nvidia's Cosmos world model for robot control and Meta's V-JEPA 2, which predicts physical world changes to help robots grasp objects in unfamiliar environments. WALL-SS takes as action input only the gripper's position, its orientation and whether it is open or closed, which leaves fine contact tasks and multi-finger dexterous hands unvalidated. |
|
| Ant Group and investment bank CICC tune a model for equity research Ant Group's Bailing team released Ling-3.0-flash-Fin, a finance-tuned large language model built with the investment bank CICC. What was added is continued pretraining on financial text and post-training on the domain rather than a new architecture, the cheaper route to a specialist model. The model keeps Ling-3.0-flash's design, at 124 billion parameters with 5.1 billion of them activated on each token. Ant is targeting four capabilities: information retrieval, research reasoning, valuation modeling and research-report writing, roughly the sequence an equity analyst works through. Its general capability score on the Artificial Analysis Intelligence Index, a composite of public benchmark results, rose from 38 to 41, a score for broad ability rather than for the finance work. The model gets a month of free access through the OpenRouter marketplace, and Ant says the weights will be open-sourced next week. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Circuit board maker Shengyi plans a $336 million plant for AI server boards Shengyi Electronics says a wholly owned subsidiary will build a high-speed interconnect circuit board plant in Ji'an, a project it calls Chip Compute Smart Connect. Boards of this class sit between an AI accelerator and the rest of the server rack, a supply layer that draws far less attention than the chips it connects. High-density interconnect boards pack their wiring more tightly than ordinary circuit boards, routing signals between chips inside a server or a car. Shengyi is aiming the capacity at AI servers and automotive electronics, with planned investment of about 2.257 billion yuan ($336 million). Designed annual capacity is 445,900 square meters of finished board, output measured by area rather than by unit count because board sizes vary. |
|
| IV | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Humanoid maker Leju trains a control model on one robot body for factory work Leju Robotics is launching KUAVO VLA, a robot control model trained on more than 600 hours of data from a single robot body, its KUAVO humanoid. Leju's premise is that a general-purpose robot foundation model does not survive a factory floor, and that the missing layer is a training pass on one body in one domain. A vision-language-action model turns camera images and a written instruction directly into robot motion. The model ships with roughly a hundred industrial skills built in, among them sorting, handling and machine loading. In Leju's own evaluation of mostly custom industrial tasks, KUAVO VLA scored 48% on task success against 16% for the general-purpose base model Lingbot-VLA 2.0. Executive Vice President Ke Zhendong said the model and its tool chain are aimed at third-party developers building industrial skills on top of it. |
| ACE Robotics and the University of Hong Kong give robot control models a short-term memory ACE Robotics and the University of Hong Kong released StreamPI, which gives a robot control model a memory of what it just saw. Most deployed control models decide from a single camera frame, which is why a robot loses track of an object it saw moved a second earlier. StreamPI restructures attention on pi-0.5, an open control model, and adds no parameters to it. It binds the task instruction to every observation so the goal does not drift as frames accumulate, and holds encoded history in cache so each step processes only the newest frame. Across four real-robot tasks, success on a cup-swap shell game rose from 47% to 80%, with grasping rolling objects improving by a similar margin. On LIBERO, a standard manipulation benchmark, the long-horizon split rose from 92.4% to 95%. |
| Robotics startup Kunlunx AI places second on a world model benchmark with a three tower design Kunlunx AI's GeWu model placed second overall on Track 1 of WorldArena 2.0 and won the image quality category outright. Track 1 scores whether what a model generates is physically and causally plausible, which decides whether a control policy tested in software behaves the same way on hardware. GeWu splits into three Transformer towers with separate parameters, one encoding the language instruction and the scene, one modeling the action sequence and one generating the future video frames. A mechanism the company calls united causal attention enforces one-way visibility between them, so the action tower can see the intent tower but not the frames it will produce. In ablation tests, cutting either causal pathway made the model degrade into reproducing the nearest training clip, with objects moving without any corresponding action. |
|
| V | QUANTUM & FRONTIER SCIENCE | |
| Chinese astronauts repair a Tiangong solar wing and add debris shielding Zhu Yangzhu and Zhang Zhiyuan repaired a solar wing and installed debris shielding outside the Tiangong space station, closing a spacewalk of about five and a half hours. It was the second time Chinese astronauts have repaired a station solar wing from outside the vehicle, the maintenance skill that decides how long a station stays in service. A solar wing is the deployable panel array that supplies the station's power. Li Jiaying supported the pair inside, backed by ground research staff. The shielding guards the station's exterior against small orbital fragments, the routine hazard of an orbit crowded with spent hardware. The three arrived aboard Shenzhou-23 in late May and have now worked more than three months on orbit. |
|
| SenseTime says domestic chips served Zhipu's new model at Nvidia-level cost SenseTime has named itself the supplier of the domestic compute behind Zhipu's GLM-5.3-Flash (covered yesterday) and says the platform was its own SenseCore. The technique it credits is heterogeneous mixed inference, which spreads one serving workload across chips from different vendors, so a fleet of several suppliers' parts runs as one. In its anonymous pre-release run as Ox-Alpha, the model served 62 trillion tokens on Chinese chips, SenseTime says. SenseTime puts price-performance for the mixed approach at 1.25 times that of Nvidia's H-series accelerators, and per-token cost at parity with mainstream Nvidia GPUs. QbitAI carried the figures under a sponsored byline. SenseTime says its Token Factory is targeting more than 10 trillion tokens served a day by the end of 2026. |
| China Unicom switches on free BeiDou satellite messaging in Tibet disaster zone China Unicom said customers in the disaster-hit area of Gyirong county, in Tibet's Shigatse prefecture, will get free BeiDou satellite messaging and free service continuity. BeiDou is China's satellite navigation system, and unlike a pure positioning service it also carries short text messages in both directions, so the constellation doubles as a fallback channel when towers are down. A BeiDou-capable terminal passes messages by satellite for disaster reporting, rescue coordination and safety check-ins. The disaster is a debris flow that began on the Nepal side of the border and struck the Gyirong port crossing, causing heavy casualties and leaving people missing. The National Commission for Disaster Prevention, Reduction and Relief and the Ministry of Emergency Management have activated a Level II relief response, the second-highest tier. |
|
|
|