China: What I Learned from the AI Labs, Robotics Startups and Academia
5 hours ago
- #Open Source AI
- #Technology Ecosystems
- #AI Labs in China
- Matt White spent eight days in April 2026 visiting AI labs, robotics startups, and academia in Beijing, Shanghai, and Hangzhou, based on direct meetings.
- Five major Chinese AI models were released during the week, including Kimi K2.6, Qwen 3.6-27B, MiMo V2.5, Ling-2.6-1T, and DeepSeek V4.
- Policy shifts occurred, with the US proposing restrictions on AI model distillation and China blocking Meta's acquisition of Manus and limiting US investment in Chinese AI firms.
- Chinese AI labs exhibit a culture of humility, focus on open source as default, and view AI as empowerment rather than job replacement, with young, lean teams.
- DeepSeek is highly respected for efficiency innovations, Moonshot AI shows rapid growth with Kimi, and Z.ai navigates geopolitical challenges with technical excellence.
- Alibaba's Qwen models dominate open-source downloads globally, and ByteDance Seed leads in video generation with strong infrastructure and data advantages.
- Chinese robotics, led by companies like Unitree, is expanding rapidly with government support, focusing on humanoid robots and manipulation challenges.
- Open source is central in China, but labs adopt hybrid models (open weights with paid APIs) for sustainability, with gaps in upstream contributions and model licenses.
- Chinese AI economics are shaped by compute constraints driving architectural innovation, domestic chip development, and a lack of a robust data economy.
- Recommendations include engaging Chinese contributors in open source, improving model licenses, and fostering collaboration to prevent fragmentation in AI and robotics ecosystems.