At the Baidu AI Developer Conference on April 16th, Baidu's founder, chairman, and CEO, Robin Li, revealed the progress of WENXIN YIYAN (文心一言) after its one-year release: the user base has exceeded 200 million, with over 200 million daily API calls, and serving approximately 85,000 customers or enterprises. Over the course of a year, the inference cost of the WENXIN large model has decreased to just 1% of its original value. Baidu's daily code additions now see 27% generated by Comate, an intelligent code assistant based on the WENXIN large model.
Li also indicated that in the future, most large-scale AI-native applications will be based on MoE (Mixture of Experts) models, meaning a mix of different sizes will be used, tailored to specific scenarios. In certain contexts, finely-tuned smaller models can achieve the effectiveness of larger ones. Recently, Li mentioned internally that whether in China or the United States, the strongest foundational models are closed-source, and various excellent smaller models are distilled from large models. Models derived from large models through dimensionality reduction are inherently superior.
