|
|
|
|
Bottom Line Up Front Alibaba releases Qwen3.8, a 2.4-trillion-parameter foundation model published Aug. 3 with substantial gains in coding and professional office work, available through the Qwen AI platform and built into the Qwen Office agent launched the same day. It ranks second behind Anthropic's Claude series on the Arena leaderboard. Data center power up 44%, reaching 49.4 billion kWh in the first half of 2026 per National Energy Administration figures, with data centers and EV charging adding 0.9 percentage points to national electricity consumption growth. V4-Flash costs 3 cents a task, about a hundredth of the $3.15 Artificial Analysis estimates for Anthropic's Claude Fable 5 on the same measure, against about 86 cents for Moonshot's Kimi K3 and $1.86 for OpenAI's GPT-5.6 Sol. |
| Alibaba releases a 2.4-trillion-parameter Qwen3.8 and will open-source its Max version next week Alibaba released Qwen3.8, the newest generation of its foundation model, on Aug. 3, and said it expects to open-source the Qwen3.8-Max version next week along with a smaller Qwen3.8-27B. The model has 2.4 trillion total parameters, the adjustable values a model learns in training. Alibaba says coding and professional office work, which it brands Cowork, improved substantially over the previous generation. Qwen3.8 is available to developers through the company's Qwen AI platform, and the model is built into Qwen Office, an agent product Alibaba launched the same day. An agent is software that carries out multi step tasks rather than only answering questions. On the third party Arena leaderboard, Alibaba's Qwen models ranked second behind Anthropic's Claude series. |
| Data center electricity use rose 44% in the first half, with computing headed for 6% of China's power by 2030 Internet data services, the category that covers data centers, used 49.4 billion kWh of electricity in the first half of 2026, up 44% from a year earlier, per National Energy Administration data. Data centers and electric vehicle charging together lifted growth in nationwide electricity consumption by 0.9 percentage points. Jiang Yi, spokesman for the National Development and Reform Commission, China's top economic planning body, said on July 31 that AI related industries grew more than 30% in the first half. National Data Administration figures put average daily token calls nationwide at 140 trillion in March 2026, against 100 billion in early 2024. A token is the text or image fragment a model processes, so the count tracks how much AI is actually being run. Computing centers used 170 billion kWh in 2025, and projected 2030 consumption is 800 billion kWh, about 6% of nationwide use. |
DeepSeek's V4-Flash runs a task for about 3 cents, a hundredth of Claude Fable 5's costArtificial Analysis, a firm that benchmarks AI models, estimates DeepSeek's V4-Flash, which entered public beta as CTD reported in its July 31st edition, costs about 3 cents per completed task on average. It puts Anthropic's Claude Fable 5 at about $3.15 on the same measure, roughly 105 times as much. Cost per task counts every token a model consumes to finish a job, which the firm says is more meaningful than headline per token pricing. Moonshot AI's Kimi K3 costs about 86 cents per task on that measure and OpenAI's GPT-5.6 Sol about $1.86. Artificial Analysis prices V4-Flash at $0.14 per million input tokens and $0.28 per million output tokens. |
| Memory maker CXMT nears mass production of LPDDR6 CXMT's LPDDR6 memory is nearing the end of research and development validation, the second of the four stages that run from die testing to formal mass production, industry figures told Yicai on Aug. 1. CXMT, formally ChangXin Memory, makes dynamic random access memory, or DRAM, the working memory in phones and servers. LPDDR6 is the newest low power DRAM standard for mobile devices, one generation beyond the LPDDR5X parts now shipping. The first CXMT part is specified at a data transfer rate of 12,800 megabits per second and holds 16 gigabits per die. Market reports in March said CXMT had sampled LPDDR6 to core customers, with a mass production ramp expected in the second half of 2026. SK Hynix said it had completed development validation of a 16 gigabit LPDDR6 first, and that supply would start in the second half of 2026. |
Carnegie Mellon adviser says Moonshot AI founder Yang Zhilin turned down Apple to build a company in ChinaRuss Salakhutdinov, a Carnegie Mellon University computer science professor and former head of AI research at Apple, said in an Aug. 1 interview with National Business Daily that he supervised Yang Zhilin's doctorate. Yang declined an Apple approach and returned to China to run Recurrent.AI, the company he had co-founded in 2016, then started Moonshot AI in 2023, whose Kimi K3 model CTD covered in its July 30th edition. Salakhutdinov said Yang told him, "if I don't try starting a company, I will definitely regret it." His congratulatory post on X about K3 set off a debate in the United States over why researchers like Yang did not stay. Yang was a core contributor to the Transformer-XL and XLNet research papers during internships at Google Brain and Meta, and each has been cited more than 20,000 times. Salakhutdinov said K3's most distinctive advantage is its open weights. |
|
| Huawei open-sources a memory layer that lets AI agents carry context between sessions Huawei's Noah's Ark Lab, the company's AI research arm, open-sourced MindMemOS, a transferable memory layer for AI agents. The design stores user preferences, project context and accumulated experience so they survive a switch between agents or a new session. Memory here means stored facts a system retrieves later, not the model's own parameters. An offline process the lab calls Dreaming finds duplicate, conflicting or superseded entries and archives outdated versions. The lab's own tests score MindMemOS at 94.03 on LoCoMo, a test of recall across long conversations, and 70.63% on PersonaMem, which tests personalization. It reports that Dreaming compressed 19.4% to 23.5% of active memory while raising question answering accuracy by up to 10.3 percentage points. |
| SenseTime open-sources an 8-billion-parameter image model with 4K output SenseTime, a Chinese AI company built around computer vision, said on Aug. 3 that it had open-sourced a preview version of SenseNova U1.5-Lite-Preview. The model is a unified multimodal system, meaning one model that both interprets and generates images and text. Its 8 billion parameter scale is small next to frontier systems, which run to hundreds of billions of parameters or more. Against its predecessor U1, the company says the new version adds native 4K image generation, finer textures and more accurate Chinese and English text rendering inside images. SenseTime says the preview significantly surpasses U1 on mainstream generation and editing benchmarks, and that its output at that scale compares with commercial closed source models. The model is published on GitHub under the OpenSenseNova organization and on Hugging Face and ModelScope. |
Sixteen chip and cloud platforms ran MiniMax's video model on the day its weights went publicMiniMax open-sourced the weights of its H3 multimodal generation model on Aug. 3, following the model's release on July 31, which CTD covered in its July 31st edition. Sixteen chip makers, developer communities and cloud inference platforms completed day zero adaptation, meaning H3 ran on their hardware and services from the day the weights appeared. Huawei's Ascend accelerators, domestic graphics chip designer Moore Threads and AMD are among them. Weights are a model's trained parameters, so publishing them lets anyone run or modify the system on their own hardware. H3 generates video up to 2K resolution and 15 seconds long with native stereo audio, and its base module deploys through inference frameworks including SGLang and vLLM. A high compression tokenizer cuts the number of tokens each video consumes, which holds down generation cost, per 36Kr. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Horizon Robotics holds a bigger share than Nvidia in chips for Chinese brand assisted driving Horizon Robotics held 31.94% of the intelligent driving chip market in Chinese brand passenger vehicles in the first half of 2026, per Gaogong Intelligent Vehicle Research Institute data. Nvidia followed at 29.38%. Horizon is a Chinese designer of chips for driver assistance systems, and intelligent driving here means assisted driving with a supervising driver, not full autonomy. City navigate on autopilot, the higher tier, handles routed driving on urban streets under supervision. In that tier Nvidia still led at 39.43%, while Horizon rose to 22.82% on mass production of its Journey 6 chips. Horizon, Nvidia and Huawei together held nearly 80% of that tier. |
|
| IV | ROBOTICS & AUTONOMOUS SYSTEMS | |
AgiBot and Unitree built most of China's first-half humanoid robots, with Unitree's sales still led by labs and schoolsThe two firms formed a duopoly in first half 2026 humanoid robot output, per MIR DATABANK data. AgiBot said 15,000 units had come off its assembly line by early June, against six prototypes in 2023 and 5,000 units at the end of 2025. Unitree completed about 11,000 units of mass production by March. Through the first three quarters of 2025, more than 70% of Unitree's humanoid robot revenue came from research and education customers, with industrial applications at about 9%. UBTech ranked third in production but fourth in shipments, which the analysis says usually signals inventory buildup or delivery lag. Unitree plans to raise annual capacity to 30,000 units. Asia Pacific director Irving Chen said shipments this year should at least double 2025's roughly 6,500 units, as CTD covered in its July 28th edition. |
| WeRide plans Denmark's first commercial robotaxi service for the first half of 2027 WeRide and GreenMobility announced a partnership on Aug. 3 to build Denmark's first commercial autonomous shared mobility project. WeRide operates driverless ride hailing fleets, and GreenMobility runs a Danish car sharing network of more than 1,500 electric vehicles. The two plan to launch public commercial robotaxi service in the first half of 2027, subject to regulatory approval. They will use WeRide's latest generation GXR, which the company describes as compliant with EU requirements. The GXR is a Level 4 vehicle, meaning it drives itself within a defined area with no human backup. Denmark is the sixth European country WeRide has entered, after it announced service for Madrid and Zurich in June. |
|
| State Council raises infringement damages for chip layout designs, effective Oct. 15 Premier Li Qiang signed a State Council decree promulgating the revised Regulations on the Protection of Integrated Circuit Layout Designs, per Xinhua News Agency. The State Council is China's cabinet, and the regulations govern how chip designs are registered in China and what protection they carry. A layout design is the physical arrangement of circuit elements on a chip, the blueprint a foundry builds from. It expands the scope of protection, clarifies how the boundaries of an exclusive right are set and increases compensation for infringement. The text also states that layout design protection work must carry out Party and state intellectual property strategy. |
| Xinhua affiliated daily says U.S. industry is turning toward open weights, naming three Chinese models Economic Information Daily, an outlet affiliated with the Xinhua News Agency, said U.S. industry is shifting from chasing frontier closed source model capability toward open models and cost effectiveness. Sina republished it on Aug. 3. It names DeepSeek, Moonshot's Kimi and Zhipu's GLM as Chinese models narrowing the gap with leading U.S. systems. It points to a joint statement from Nvidia, Microsoft and IBM, among others, urging U.S. policymakers to support open weight AI models, meaning models whose trained parameters are published for anyone to run. It also notes the Open Secure AI Alliance that Nvidia and other companies announced on July 27 to develop and share open source security tools. It cites Associated Press reporting that experts consider the latest Zhipu and Moonshot models nearly as intelligent as the frontier models from OpenAI and Anthropic, at significantly lower cost. |
|
|
|