China AI Native Industry Insights – 20260708 – ByteDance | Alibaba | Tencent | more
Explore EdgeBench, Alibaba’s RAG service, and Tencent’s Hunyuan Hy3. Discover more in Today’s China AI Native Industry Insights.
1. ByteDance releases EdgeBench benchmark for evaluating AI agent learning in real-world environments and discovers new scaling law
ByteDance released EdgeBench, a benchmark designed to evaluate AI agent learning in real-world environments, featuring 134 real-world tasks across six capability domains with continuous interaction periods of at least 12 hours per task. Based on approximately 38,000 hours of agent-environment interaction data, researchers discovered that agent performance during environmental learning follows a high-precision log-sigmoid curve with an average R-squared fit of 0.998. The study also found that learning speed of frontier models roughly doubles every three months across recent generations. The project has open-sourced 51 tasks and the complete evaluation framework on GitHub and Hugging Face for community research.
Read more: https://mp.weixin.qq.com/s/TRdasyCjWlExZTilsYi_bg
Video Credit: NotebookLM
2. Alibaba Cloud Bailian launches enterprise-grade Agentic RAG service with multi-library retrieval and multimodal capabilities
Alibaba Cloud Bailian has launched Knowledge Studio, an enterprise-grade Agentic RAG service that supports multimodal search, agentic search, and joint retrieval across up to 15 knowledge bases. The service features six built-in retrieval tools including semantic search, chapter browsing, and SQL execution, with automatic fallback strategies when initial retrieval fails. The Agentic Generation capability can cite sources across multiple modalities including images, tables, and documents with traceable references to original documents, paragraphs, and page numbers. Companies like Geely Automotive and AIA have adopted the platform for applications ranging from R&D efficiency to 24/7 insurance consultation services.
Read more: https://mp.weixin.qq.com/s/oZaEP2CjtmxSpVMhXzqJIA
Video Credit: The original article
3. Tencent releases Hunyuan Hy3 large language model with improved agent capabilities and integration across 50+ business products
Tencent officially released Hunyuan Hy3, an upgraded version of its large language model following the preview released in late April. The model features 295B total parameters with 21B activated parameters using MoE architecture and supports up to 256K context length. Hy3 has been integrated into core Tencent products including WorkBuddy, CodeBuddy, Yuanbao, WeChat Official Accounts, QQ Browser, and WeGame, with nearly 50 additional business lines in the integration queue. The model demonstrates significant improvements in agent tasks, with WorkBuddy showing task success rates increasing from 72% to 90% and average processing time reduced by 34%. Hy3 is available on Tencent Cloud with pricing of 2 yuan per million input tokens and 8 yuan per million output tokens, and is open-sourced under Apache 2.0 license for global developers on platforms including Hugging Face and ModelScope.
Read more: https://mp.weixin.qq.com/s/9ntGUi3KqtbSn-Ngvv7jmg
Video Credit: The original article
That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.