Baidu CTO Wang Haifeng Highlights Wenxin Yiyan Large Language Model at Zhongguancun Forum

From May 25 to 30, the Zhongguancun Forum was held in Beijing, jointly organized by the Ministry of Science and Technology, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the State-owned Assets Supervision and Administration Commission, the Chinese Academy of Sciences, the Chinese Academy of Engineering, the China Association for Science and Technology, and the Beijing Municipal Government. Over 80 countries and regions participated, focusing on frontier fields such as artificial intelligence, life sciences, and clean energy. Baidu’s Chief Technology Officer Wang Haifeng attended the AI Large‑Model Development Forum and delivered a speech titled “Wenxin Yiyan, Reading Ten Trillion Books,” explaining the technology and industry model of large language models and showcasing Wenxin Yiyan’s capabilities in knowledge Q&A, document writing, poetry creation, video generation, reasoning, and intelligent summarization in work scenarios, while envisioning deep integration of large models with the real economy.

The Chinese leadership has emphasized the importance of self‑reliant scientific and technological development. The Politburo meeting on April 28 called for strengthening the foundation of technological self‑reliance, fostering new growth drivers, and creating an innovative ecosystem while managing risks. The 20th Central Financial and Economic Committee highlighted the need to harness new technological revolutions such as AI, to promote industrial intelligence, greening, and integration, and to build a modern industrial system that is complete, advanced, and secure.

Large models built on strong algorithms, massive compute, and big data have become the mainstream direction of AI development. Models that can understand, reason, create, and even exhibit emotional traits are seen as a beacon for achieving general artificial intelligence.

Wenxin Yiyan is Baidu’s self‑developed knowledge‑enhanced large language model, based on the Wenxin series and pre‑trained on more than a trillion pieces of data and knowledge. It employs supervised fine‑tuning, reinforcement learning from human feedback, and prompting techniques, and incorporates knowledge‑enhancement, retrieval‑enhancement, and dialogue‑enhancement technologies. Benefiting from Baidu’s comprehensive AI stack and the PaddlePaddle deep‑learning platform, Wenxin Yiyan’s inference performance improved nearly tenfold within a month, delivering higher efficiency and better results. PaddlePaddle now hosts 5.35 million developers, serves 200 thousand enterprises, and has enabled the creation of 670 thousand models, leading the Chinese deep‑learning platform market.

In his presentation, Wang demonstrated Wenxin Yiyan’s multi‑scenario abilities. When asked about the major scientific achievements announced at the 2023 Zhongguancun Forum opening ceremony, the model listed ten key breakthroughs with high timeliness and accuracy, showcasing its retrieval‑enhancement advantage. For questions on the forum’s significance for China’s independent innovation, the model provided a clear, well‑structured summary. It also composed an acrostic poem titled “Innovation Development” and generated a video introducing the forum.

Within Baidu’s internal workflow, Wenxin Yiyan serves as an “intelligent office super‑assistant,” integrated with platforms such as the Intelligent Work Platform (Liu). It enables human‑AI collaboration by providing instant IM chat summarization, code generation, and knowledge‑base creation, acting as a knowledgeable, real‑time assistant that boosts employee productivity.

AI is now a major driver of the new wave of scientific and industrial transformation. Large models offer strong performance, broad generalization, and standardized development processes, creating new opportunities for the AI industry. To address industrialization challenges, Wang likened the emerging model‑production ecosystem to a chip foundry model: enterprises that combine algorithms, compute, and data can encapsulate the complex model‑building process into a low‑threshold, high‑efficiency production platform serving thousands of industries. This approach has already been validated in the Wenxin model practice, with Baidu co‑building industry‑specific models with leading partners in energy, finance, aerospace, manufacturing, media, urban management, social sciences, and film.

Looking ahead, large models will reshape cloud computing. Model‑as‑a‑Service (MaaS) will become the mainstream cloud business model, with applications built on large models across all sectors. Each industry will need its own tailored model, and the deep integration of large models with the real economy will empower countless industries, accelerate transformation, and promote high‑quality economic development.