The Beijing municipal government released a policy titled Several Measures on Accelerating the Development of Agent AI in Beijing on Thursday, one of the first city-level frameworks worldwide to target AI agents as a distinct category from generic AI systems, per Pandaily. The document spans foundation model capability, developer tooling it labels Harness Engineering, skill marketplaces, an AI operating system layer and infrastructure for what officials call the Token economy, including planned Token factories. It reorients evaluation away from static benchmarks toward reliability on multi-step tasks, citing industry estimates that even a 99.9% per-step accuracy still leaves a 100-step agent workflow failing roughly once in 10. The remaining measures cover computing resource pooling, industry vertical applications, talent development, international cooperation and governance. The plan embeds agent-specific rules into Chinese industrial policy at the municipal level, layering on top of national agent guidelines that China's cyberspace, planning and industry regulators jointly issued earlier this year.
Read at Pandaily ↗
DeepSeek founder Liang Wenfeng told investors in a circulated four-hour transcript that the lab prices its API to recover hardware costs in about 10 months rather than maximize profit, per TechNode. Liang said DeepSeek plans to keep its strongest models open source and identical to what it runs internally. He said DeepSeek is not trying to become "the next ByteDance or Tencent" and views consumer traffic, enterprise revenue and API pricing as secondary to its AGI roadmap. Liang framed the technical path as chain-of-thought reasoning, then agents, then continual learning and eventually embodied intelligence, with coding agents prioritized because they can improve both products and internal research. DeepSeek is not prioritizing 3D generation, video generation or world models, which Liang views as useful applications but not central to intelligence development.
Read at TechNode ↗
iQIYI on Thursday exclusively released "Qitan: Paper Blade Across the Wasteland," a 60-plus-minute AI generated online story film co-produced with DaMeng Innovation that the company said is the first AIGC online story film in China to obtain a Network Drama and Film Distribution License and go live under it, per ITHome. A team of more than 20, including prompt engineers and AI asset effects artists alongside conventional film roles, built a full-process pipeline spanning creative development, prompt workflows, visual asset generation, AI redrawing, 4K output and postproduction. The team drew on classical texts including the "Classic of Mountains and Seas," "Zi Bu Yu" and "Liaozhai" to complete more than 200 character and visual asset designs. The film opens a planned "Qitan" trilogy, directed by Zeng Xiangcheng with Liu Haiyang credited as AI director; iQIYI separately announced in April a fully AI generated film based on its "Soul Ferry" series, due out in 2026.
Read at ITHome ↗
Chinese AI models narrowed the performance gap with U.S. counterparts to a record-low 6% in June, down from 9% in May, which Bloomberg Intelligence said raises questions about whether U.S. technological supremacy in the sector is sustainable, per Bloomberg. Chinese models took two slots in LiveBench's global ranking of large language model performance last month. The gap had averaged about 10% to 15% over the prior 12 months before Zhipu AI's GLM-5.2 model reached the top of the global ranking in agentic coding.
Read at Bloomberg ↗