Open-Weight Models From China Are Set to Supply BRICS Countries With AI

At the BRICS summit in New Delhi, Xi Jinping offered the group’s 11 members help building a shared open-source AI base. The timing was awkward for Washington: in the U.S., the argument is not about sharing models, but about whether the most capable ones should be slowed down at all.

China has something concrete to put on the table. It leads the open-weight pack with names like Moonshot AI, DeepSeek, Z.ai, MiniMax, Xiaomi and Alibaba. Moonshot’s Kimi K3 is described as the largest open-weight model released so far, while Alibaba’s Qwen family has become a default starting point for countless fine-tunes around the world. Some of these models are already being trained on Huawei chips instead of Nvidia hardware.

That matters for countries without giant data centers. If you cannot build a gigawatt-scale AI stack, downloadable weights are the quickest way to get one off the ground. Xi’s pitch is aimed squarely at that gap, and he wrapped it in a broader BRICS offer: trade and investment facilitation, a cloud platform, cooperation on smart factories and manufacturing standards, and an engineer training alliance.

His speech leaned hard on rule-setting. Xi said international rules should be written by all countries, not just the powerful ones, and called for WTO reform, a bigger role for developing countries in the international financial system, and a consensus-based global AI framework that can support shared prosperity.

But the open part of the offer is still fuzzy. China’s Ministry of Commerce is reportedly considering export controls on AI models, including filing rules for weaker ones, security reviews for stronger ones, and even a ban on public release for the most capable systems. The government has not decided, and talks with Alibaba, ByteDance and Zhipu are still ongoing. So the BRICS open-source community exists for now mostly as an idea, with funding, timing and model selection all unresolved.