中国 AI 周报：智谱认领神秘模型、DeepSeek 估值接近 $74B、北京举办机器人奥运会

China AI Weekly Digest summarizes the week’s big news, model updates, policy, products, and research, sourced from the aggregated newsfeed of China AI Index. The digest below is mostly compiled and drafted by an AI agent sourcing from credible news outlets, then fact-checked and edited by me.

This week: 322 stories tracked (5 editor picks, 112 English-language, 205 Chinese-language 🇨🇳) over 2026-08-22–2026-08-29.

The Big Three

A mystery model that quietly topped the charts turns out to be Zhipu’s, running entirely on domestic Chinese chips. An anonymous model calling itself “Ox Alpha” had developers guessing for six days. Some speculated it was Cursor trained on open GLM weights. Then Zhipu (Z.ai) claimed it as GLM-5.3-Flash: a natively multimodal, 320-billion-parameter (18B active) model running inference entirely on domestic Chinese chips, priced at roughly 1/40 of Claude Opus 4.8. Z.ai’s shares jumped on the reveal, and two days later the company open-weighted its larger flagship, GLM-5.3 (744B total/40B active), under an MIT license. (SCMP/Bloomberg/CNBC/@Zai_org)

DeepSeek edges toward a 2027 listing as revenue rockets and its valuation nears $74 billion. DeepSeek is closing in on a pre-IPO funding round as it lines up a public-market debut as early as 2027, with the Wall Street Journal reporting a target valuation of $74 billion; The Information separately reported DeepSeek’s revenue reached $70 million in July, a tenfold jump from a year earlier. The fundraising push doubles as a coming-out moment for founder Liang Wenfeng’s quant fund High-Flyer, which bankrolled DeepSeek from the start and is now navigating China’s crowded, choppy IPO market on its behalf. (WSJ/CNBC/SCMP/The Information)

A robot beats Usain Bolt’s 100m record at China’s World Humanoid Robot Games. Beijing hosted the second edition of the World Humanoid Robot Games, a five-day, 51-event spectacle spanning running, table tennis, soccer and boxing. The event has become a showcase for China’s robotics progress. This year’s edition featured a robot beating Usain Bolt’s own 100m sprint record. Homegrown humanoid maker AGIBOT dominated the games, topping both the gold-medal and overall-medal tables with 46 medals in total (18 gold, 16 silver, 12 bronze). (AP/BBC/@AGIBOTofficial)

A wave of new releases crowded the week’s model news.

Tencent’s Hunyuan team previewed Hy4, featuring 770B parameters, 49B active, a 1M-token context window. Tencent says the model outperforms both Z.ai and Moonshot. (Bloomberg/@TencentHunyuan); Alibaba’s Qwen team launched Qwen3.8-Flash, a smaller multimodal MoE model billed as an early preview of the Qwen4 architecture, cutting training and serving costs. (Bloomberg/@Alibaba_Qwen); MiniMax said it has begun merging its M3 and H3 models into a single multimodal model and expanded its Alibaba Cloud compute partnership to keep pace with training and inference demand. (CNBC/SCMP); Xiaomi is developing a new self-designed “Xring” chip and has lined up TSMC to manufacture it, sources told Reuters. (Reuters); Ant Group’s Ling team shipped Ling-3.0-flash-Fin, built for financial-sector use with financial institutions and domain experts. (@AntLingAGI)

Funding

Money kept moving fast this week into AI labs, and out of some newly public robotics stocks.

The Guardian reports China’s government has grown wary that AI companion apps could substitute for human intimacy, even as talking to AI has become mainstream there. (Guardian); Brazil’s data-protection authority fined ByteDance over unlawful processing of teenagers’ data on TikTok. (Reuters/财新)

ByteDance is folding its AI tools, TRAE and Coze (扣子), into Doubao, building a unified “Doubao Work” brand to take on Tencent’s Workbuddy and Alibaba’s Qwen Work. (Bloomberg/36Kr); DeepSeek’s family of cheap, open-weight models is leading a broader surge of Chinese models onto a major US model-hosting platform. (SCMP); Baidu’s ERNIE-powered assistant rolled out upgrades at Baidu AI Day, featuring 400+ professional skills across 30+ industries, with daily active users up 83% year-over-year in June. (@Baidu_Inc)

Security researchers say Chinese hackers are using DeepSeek to help scale their attacks, Bloomberg reports. (Bloomberg); Meituan’s LongCat team evaluated seven frontier models on 36 AI R&D tasks to test whether agents can act as autonomous researchers. (@Meituan_LongCat)