Beijing Trains Global South Devs on Free Open AI Models

The interesting piece of China's AI push at this year's World AI Conference in Shanghai isn't another leaderboard-topping open model, it's the training pipeline underneath it. As the Financial Times reports, Xi Jinping used the July 2026 gathering to promise 5,000 AI training and seminar slots for developing countries over the next five years, alongside cooperation centres tied to ASEAN, the African Union and BRICS. Representatives of 29 countries also signed on to a new intergovernmental body, the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai.

Capacity-building sits alongside the model releases, not behind them. Alibaba representatives have reportedly been offering developers free training and free tokens on Qwen; DeepSeek and Moonshot's Kimi keep shipping weights cheap enough to fine-tune on modest hardware. The adoption numbers are already visible. Alibaba says Qwen passed one billion cumulative downloads on Hugging Face by January 21, 2026, and by March it was capturing more than half of global open-source model downloads. Four years ago US models were roughly 60 percent of Hugging Face downloads; by November 2025 that share was down to 16 percent.

Why this matters if you are not shipping models yourself: the sovereign AI deals now being written across the Global South are being written on top of whatever stack local developers were trained on. Singapore's government-backed AI Singapore program chose Qwen over Meta's Llama for its latest regional model. Malaysia said its sovereign AI ecosystem would run on DeepSeek. In Uganda, Sunbird AI's Ernest Mwebaze built Sunflower, an LLM covering 31 Ugandan languages, on Qwen 3 rather than a Google, Microsoft or OpenAI base. Zhipu AI, which is backed by the Chinese Communist Party, is reportedly selling into governments and state-owned enterprises in Malaysia, Singapore, the UAE, Saudi Arabia and Kenya.

Take the specifics as reported, not as a settled scorecard. What the coverage doesn't give you is a delivery breakdown of those 5,000 training slots, who runs them or where the graduates end up, and it doesn't say what the WAICO body can actually enforce on its 29 signatories. It also doesn't separate wins that are genuine commercial preference from those helped along by subsidy or diplomacy.

Still, the direction is the part worth watching. If the developer habit layer of the next few years belongs to Qwen, DeepSeek and Kimi, closed-API vendors in the US don't just lose downloads, they lose the shape of the government contracts written on top of them.