DeepSeek’s open source models accelerated by Baidu’s Baige 5.0 platform
Baidu has launched its Baige 5.0 AI infrastructure platform in Beijing, built to accelerate DeepSeek’s open-source models and strengthen China’s domestic AI technology stack. The system is powered by a mix of semiconductors, including chips designed by Baidu’s Kunlunxin unit, to improve model training and inferencing efficiency.
The upgraded platform delivers faster network speeds, greater computing power, and enhanced performance through adaptive and smart resource allocation that reduces latency and increases throughput. According to Baidu executive vice-president Shen Dou, who also serves as president of Baidu AI Cloud Group, DeepSeek’s R1 reasoning model achieved around a 50 per cent improvement in inferencing efficiency on Baige 5.0.
“That means, with the same time and cost … we could have the model ‘think’ 50 per cent more [or] work 50 per cent more,” Shen said at the launch.
The platform supports China’s broader effort to reduce reliance on foreign technologies and push ahead with a self-reliant AI ecosystem amid ongoing US trade restrictions. Central to this effort is the Kunlunxin Super Node, which connects hundreds of AI chips and is capable of deploying trillion-parameter AI systems within minutes. Kunlunxin aims to compete with US giant Nvidia and Chinese rivals Huawei HiSilicon and Cambricon.
Momentum in China’s chip sector is accelerating. Kunlunxin recently secured chip orders worth 1 billion yuan (US$139 million) from China Mobile, while Cambricon reported strong revenue growth in the first half of 2025 on rising AI adoption. Meanwhile, DeepSeek has hinted at the imminent release of next-generation domestic AI chips.
Despite the unveiling, Baidu’s shares fell 1.25 per cent to close at HK$86.60 on Thursday.
