China's AI labs target 'physical AI' with embodied agents

- RynnBrain's architecture is built upon Alibaba's Qwen3-VL model and uses a unified encoder-decoder to translate visual inputs and text into spatial trajectories and action plans. To encourage adoption, Alibaba has open-sourced several versions, including 2-billion and 8-billion parameter dense models and a 30-billion parameter Mixture-of-Experts (MoE) variant. - Beyond Alibaba, other major Chinese tech firms are open-sourcing agentic and multi-agent frameworks; ByteDance has released DeerFlow, a multi-agent system for automating research workflows, while Zhipu AI, a Tsinghua University spin-off, released GLM-5, a 745-billion parameter MoE model designed for autonomous agent systems and trained entirely on Huawei Ascend chips. - The push into physical AI is part of a broader national strategy to translate digital AI advancements into the "real economy." This is supported by local regulations, such as in Hangzhou, which passed China's first law to promote the embodied AI robotics industry, and Shanghai's municipal plan to create shared infrastructure for R&D, testing, and financing. - Chinese firms are projected to manufacture over 10,000 humanoid robots in 2025, which would account for more than half of the global output. This industrial-scale production is backed by a surge in investment, with China's funding for new robotics ventures in the first seven months of 2025 reaching $3.4 billion, 42% more than in the U.S. - A key commercial strategy for consumer-facing agents in China is "agentic commerce," where companies like Alibaba and Tencent are deeply integrating task-oriented AI into their existing super-app ecosystems like Taobao, Alipay, and WeChat. The goal is to create a closed-loop "full consumption journey," from product recommendation to payment, within a single chatbot interface. - The development of multi-agent systems is a defining trend, with open-source orchestration frameworks like LangGraph, AutoGen, and CrewAI becoming foundational for coordinating complex workflows between specialized agents. These frameworks manage state, route messages, and handle the handoff between agents and tools. - Recent AI agent research highlights a focus on multi-agent architectures that can be automatically generated and optimized, as well as dynamic tool retrieval where an agent selects the right tool based on the query and execution context. Survey papers on Large Language Model-based multi-agents point to challenges in aligning agent goals and ensuring collaborative reliability. - The AI agents market in China generated approximately $577 billion in revenue in 2025 and is forecast to grow at a compound annual growth rate of 50.8% through 2033. Some analysts predict that the first AI agent to surpass 300 million monthly active users could emerge as early as 2026.