Huawei has released two of its Pangu large language models under open-source licences, making a 7-billion-parameter dense model and a 72-billion-parameter mixture-of-experts version freely available alongside some of the company’s model-reasoning technology. The 1 July move gives developers worldwide the right to inspect, modify and deploy the software, marking the first time the Chinese company has opened the core of its flagship models. Analysts say the decision is aimed at pulling more workloads onto Huawei’s Ascend AI chips, which cannot easily be replaced with U.S.-made processors because of ongoing export restrictions. “The chip-to-model strategy positions Huawei across the full hardware-to-software stack,” said Ray Wang of Constellation Research, comparing the approach with Google’s pairing of its TPUs and Gemma models. The announcement came the same day Baidu put its Ernie 4.5 suite—ranging from 0.3 billion to 424 billion parameters—on open-source platform Hugging Face. TrendForce estimates that the recent wave of open-sourcing by Chinese firms, including DeepSeek and Alibaba, is cutting the cost of running large models by as much as 60 % to 80 %. The broader push is beginning to resonate outside China. A Wall Street Journal analysis citing Sensor Tower data found ChatGPT has 910 million global downloads while DeepSeek, one of China’s fastest-growing alternatives, has reached 125 million. Industry observers say the rising popularity of Chinese models is eroding the United States’ early lead in generative AI and intensifying competition for developers and cloud spending worldwide.