China's AI Models Are World Class

Wanted to weigh in on some of the debates happening in the U.S. surrounding the release of Kimi K3 and Alibaba’s Qwen3.8 Max over the last week, as well as Hugging Face relying on Zhipu’s open GLM-5.2 to analyze a recent security breach. Our teams in China have been working with open Chinese models for a while now, so this DeepSeek Moment 2.0 isn’t really surprising to us.

One overall takeaway is that the Chinese models are world class. Everyone in Silicon Valley knows that Chinese AI talent is world class, so the surprised reaction seems overwrought. The main outcome of US chip restrictions has been more efficient Chinese models that perform well on lower compute, which is actually better for users and for the environment. (Also reduced Nvidia revenue and an incentive for Huawei to build competitive products). Now Kimi K3 and Qwen3.8 Max are effectively as powerful as Fable 5. I think we need to reset the terms of the conversation.

In China we’ve been using Kimi and Qwen in our commerce-focused applications. The key advantage of these models is that by being open weight, we can download them and customize for specific tasks, and low token cost makes it feasible for us to apply the models across our services. Qwen in particular is amazing for commerce-related tasks given that it’s trained on Alibaba transaction data. Even setting aside the security guardrails issue, the open Chinese models are simply easier, cheaper, and in some cases more effective for our company to use.

Final comment—China is committed to rolling out low-cost AI applications and to achieve positive social outcomes. Both private and public sector are committed to open source. US firms seem to be positioning for a potential AGI payout, whereas in China the leading firms are focused on applications. At the World AI Conference in Shanghai last week we heard a genuine commitment to proactively addressing potential social issues arising out of AI. I think there’s more enthusiasm at the ground level among ordinary folks about AI applications—like potential use cases of humanoid robots in healthcare, as one example—rather than the anxiety we’re seeing in the US.