---
date: 2026-08-13
subject: "DeepSeek pulls V4 Pro release | Qwen3.8 runs on nine chips day one | 420 km quantum entanglement"
---

**DeepSeek pulled its V4 Pro release** hours after shipping it today, following user reports that it performed little better than the smaller V4-Flash. **FlagOS ran Alibaba's Qwen3.8 flagship** on nine AI accelerators, eight of them domestic designs, on the day the weights shipped, converting the model's FP8 weights to INT8 for chips that cannot run FP8 natively. **Chinese physicists held entanglement across 420 kilometers** of optical fiber between two cold-atom quantum memories, passing the 230 kilometers theoretical ceiling for entanglement through fiber with no relay in between.

# Top Stories

- **DeepSeek pulls its V4 Pro release hours after shipping it** — DeepSeek deleted the announcement of its V4 Pro model from its own website today, hours after the build went live. Users testing it reported performance was not much better than the smaller V4-Flash, which would put it out of reach of near-parity with Anthropic's Claude Fable 5 that DeepSeek's own benchmark scores claim. Mydrivers reports speculation that a deployment fault left the V4 Pro endpoint serving the Flash model, and that a genuine shortfall would mark down DeepSeek's financing valuation. The architecture is unchanged from April's preview, with only post-training redone. The company's agent scores moved sharply: its DeepSWE software engineering test rose from 12.8 to 62.7 on tasks involving long-chain code edits and tool calls. On knowledge reasoning and the hardest repository work the model still trailed Fable 5 and Claude Opus 4.8. The name DeepSeek-V4-Pro-0813 is still listed on the models-and-pricing page. [Mydrivers](https://m.mydrivers.com/newsview/1143495.html) [GeekPark](https://geekpark.net/news/368777)

- **A Beijing institute's porting project runs Alibaba's newest open model on nine accelerators** — The FlagOS community, a project the Beijing Academy of Artificial Intelligence organizes, had Alibaba's newly open-sourced 2.4-trillion-parameter Qwen3.8 flagship running on nine AI accelerators the day it shipped. That followed the Qwen3.8-Max version's first-place finish on an independent AI agent ranking (covered Aug. 6). One unified open-source stack carried the adaptation, precision alignment and deployment checks for all nine, eight of them domestic designs including Huawei Ascend, Kunlunxin and Hygon, instead of a separate port per chip. Alibaba shipped the weights in BF16 and FP8, a low-precision number format that cuts the memory each parameter occupies. FlagOS says more than 80% of China's installed AI compute cannot run FP8 natively. Its toolchain converts weights in either format to a unified INT8 form, and the converted model scored within tolerance of Nvidia's native version. Tencent Cloud's HAI community and China's Supercomputing Internet already carry the adapted models as ready-to-run container images. [GeekPark](https://geekpark.net/news/368778)

- **Chinese physicists hold entanglement between two quantum memories over 420 kilometers of fiber** — Two cold-atom quantum memories, devices that store a photon's quantum state and release it later on demand, held entanglement across 420 kilometers of optical fiber. Above 230 kilometers, the result passes the PLOB bound theoretical ceiling on distributing entanglement through fiber with no relay station in between. The group's scheme links entanglement using single-photon interference, so the success probability scales with half the channel's transmission efficiency rather than the whole channel. That is why the advantage appears only over long distances. Moving the memories' operating wavelength from 795 to 780 nanometers, with a different frequency conversion scheme, put the signal photons inside the fiber's lowest-loss window. Pan Jianwei's group at the University of Science and Technology of China ran the work with the Jinan Institute of Quantum Technology and a Chinese Academy of Sciences institute in Shanghai. Physical Review Letters published it Tuesday as an Editors' Suggestion. [ITHome](https://www.ithome.com/0/989/198.htm)

- **Beijing's second humanoid robot games draw 2,056 machines from 16 countries** — The second World Humanoid Robot Games open Aug. 22 at Beijing's National Speed Skating Oval, with 666 teams and 2,056 robots entered from 16 countries. That is about four times the machines the first edition drew, and Chinese companies, universities and teams account for nearly all of them. International entries include U.S., German and Japanese teams, while Brazil assembled a national side from its RoboCup championship squads. Each robot and each battery gets its own code at check-in, so team and equipment data stay traceable. A management platform for embodied intelligence, AI that operates through a physical body, then tracks joint temperature, motion state and battery level for every entrant over encrypted dedicated links. The event's training and evaluation base has run for a month, hosting hundreds of practice sessions. [ITHome](https://www.ithome.com/0/989/270.htm)

- **Automotive chipmaker SiEngine starts volume shipments of a 7nm in-car AI accelerator** — SiEngine Technology said its 7nm automotive AI accelerator has entered full mass production, with volume shipments beginning to automakers and Tier 1 suppliers. The part, branded Tiangong 100, is a plug-in accelerator rather than a new main controller. SiEngine says that lets a carmaker add in-car model inference without rebuilding the vehicle's controller and wiring layout. Its 7nm process is an advanced class, roughly two generations behind the global leading edge. The chip delivers 96 TOPS of INT8 compute, a measure of low-precision inference throughput, on a dedicated memory channel that keeps AI work off the car's base compute. That is enough to run one model of 3 billion to 7 billion parameters, or several smaller ones at once. SiEngine puts hardware cost in the thousand-yuan class, roughly $150, and says the design also extends to industrial control and robotics. [EV Industry Supply Chain](https://www.evchanye.com/2026/0813/50998.shtml)

# Policy & Regulation

- **Chinese refiners say Congo's concentrate export ban misses their Congolese output** — The Democratic Republic of the Congo issued an order banning all exports of copper and cobalt concentrates, repealing the exemption clauses in its earlier bans. Congo supplies about 70% of the world's cobalt, effecting the imports of China and other nation's battery supply chains. Huayou Cobalt, CMOC and Zijin Mining each said their Congolese output is refined material, cathode copper and cobalt hydroxide rather than the concentrate the order covers. CITIC Securities said most Chinese copper operations there use hydrometallurgy, a leaching route rather than smelting, so their cathode copper also falls outside it. An annual quota already caps Congo's cobalt metal exports for 2026 and 2027 at 96,600 tonnes, half of what it shipped in 2024. The order carries a three-month transition period, and leaves open only a strategic waiver from the mining minister. [EV Industry Supply Chain](https://www.evchanye.com/2026/0813/50930.shtml)

# AI & Foundation Models

- **Huawei says a cache-handling change cuts an AI agent's first response time by 57%** — Huawei's openJiuwen agent platform and its Ascend accelerator line released a joint technique that Huawei says cuts an agent's time to first token by more than 57%. A conventional inference engine tracks requests and cache but not what the agent is doing. Key-value cache from a finished subtask, the intermediate state a model keeps for a running conversation, can therefore sit in the accelerator's high-speed memory. That is the costliest tier. Huawei's fix is a state contract it calls Agent Hint. The agent signals key changes in a session's lifecycle, and the engine evicts, offloads or prefetches cache in response. Huawei's own measurements also put peak inference storage down 25% and the prefix cache hit rate up 33%. The company says it will open-source the capability through the openJiuwen community. [Sina Tech](https://finance.sina.com.cn/tech/discovery/2026-08-13/doc-ininamih5153319.shtml) [state-affiliated]

# Chips & Semiconductors

- **Cambricon says its accelerators now run five of China's main model families** — Cambricon has finished inference adaptation for five mainstream domestic large language models, Chairman and General Manager Chen Tianshi said at the company's interim results briefing. Adaptation is the work of getting a model to run on a given accelerator at its expected accuracy. The five named include Zhipu's GLM, DeepSeek and Alibaba's Qwen. On training, the company continues improving performance on the DeepSeek, Qwen and Hunyuan series, and has extended thousand-card and ten-thousand-card cluster runs into scenarios covering communication, fault tolerance and long-run stability. Chen said the sixth-generation processor microarchitecture and instruction set remain in development, aimed at large-model training and inference. [ITHome](https://www.ithome.com/0/989/284.htm)

# Robotics & Autonomous Systems

- **Robot maker X Square sorts 1,816 parcels an hour in a live stream test** — X Square Robot's two-armed system sorted 1,816 parcels in an hour during a public livestream yesterday. That sits above the 1,248 an hour Figure AI has published as its logistics average, on simpler hardware: two arms with a standard gripper against the U.S. developer's humanoid with a five-finger dexterous hand. Parcels arrived randomly mixed and randomly oriented, from cartons and soft mailers to foam-packed fresh goods, which QbitAI calls a harder setup than Figure AI's earlier livestream. The system runs on WALL-B, X Square's own model, which fuses vision, hearing, language, touch and action in one network rather than chaining separate perception and action modules. Handling switched by parcel: a direct grip for light items, a two-arm carry for heavy boxes and flattening garment bags before locating the label. The hour ran continuously with no human takeover, at 98% accuracy. [QbitAI](https://www.qbitai.com/2026/08/471049.html)

- **Battery maker CATL clears an aviation fire-spread test on an air-taxi pack** — CATL said its aviation battery system passed a test in which two adjacent cells were driven into thermal runaway, the self-sustaining overheating reaction inside a lithium cell, without the fire spreading. The company calls it the first 350 Wh/kg prismatic cell to meet that aviation standard. Most batteries now flying in electric vertical takeoff and landing aircraft, the air taxis known as eVTOLs, sit between 280 and 320 Wh/kg. Testers triggered runaway at both the middle and the edge of the pack, and neither position ignited surrounding cells. Representatives designated by the Civil Aviation Administration of China witnessed cell manufacturing, assembly and the full test. CATL says that lets the data feed directly into later airworthiness certification of the aircraft. The cell will fly first on AutoFlight's manned eVTOL. [EV Industry Supply Chain](https://www.evchanye.com/2026/0813/50992.shtml)

- **Chinese and Singaporean labs publish a robot model that moves while it manipulates** — Nanyang Technological University, Peking University, the Hong Kong University of Science and Technology (Guangzhou) and the Beijing Academy of Artificial Intelligence published omega-0, a robot model that completed 11 household tasks on real hardware with an 81.8% success rate in its full-scene mode. Given a language instruction, it perceives the environment and the robot's own state, then outputs compact action features a low-level controller reads directly. That supports working while moving rather than stopping to act. Predicted future visual features track task progress, while a separate branch generates whole-body commands. Tasks included stowing items across different areas, clearing clutter at high and low reach, and picking clothes off a bed, with both modes scoring above the pi-0.5, GR00T-N1.7 and psi-0 baselines. The team open-sourced omega-HOME, 40.3 hours of real home recordings. [TMTPost](https://www.tmtpost.com/8101182.html)

# Digital Infrastructure

- **Sichuan backs 10,000-chip AI clusters and compute sites built into tunnels** — Sichuan issued a trial version of its policy measures for building out a computing power network, supporting intelligent computing clusters of 10,000 or more accelerator cards. The province is courting basic telecom operators and leading large-model companies to site those clusters, and applying fiscal, financial and green-electricity tools together to land the projects. One measure directs officials to explore high-security compute facilities in tunnels and other available concealed environments, built as modular, green, high-density sites with specialized central state-owned enterprises brought in. The two zones named for the clusters are the Chengdu Plain compute core zone and the Panxi-Northwest Sichuan compute-and-electricity belt. [Wallstreetcn](https://wallstreetcn.com/livenews/3148934)