Why Baidu Is Open‑Sourcing Its Next‑Gen Ernie Model and What It Means for the AI Landscape

After announcing that its services would become free, Baidu revealed that the next generation of its Ernie (Wenxin) large language model will be open‑source, with the 4.5 series to be released gradually and the code opened on June 30.

The article reviews the evolution of Baidu’s Wenxin Yiyan (Ernie Bot) from its closed‑source launch on March 16, 2023, through its public beta on August 31, 2023, the professional subscription version on November 1, 2023, and the rapid iteration to version 4.0 within two years, highlighting its role in China’s LLM ecosystem.

From a technical perspective, Wenxin’s strengths lie in knowledge‑enhanced learning—integrating Baidu’s trillion‑scale knowledge graph to improve factual accuracy and reduce hallucinations—continual learning that updates the model without forgetting old knowledge, and multi‑task learning that shares representations across dialogue, translation, reading comprehension, and code generation. The model also supports multimodal generation of images, video, and speech, and benefits from Baidu’s ecosystem (search, maps, cloud).

Industry insight: the shift from paid, closed‑source to free, open‑source reflects broader market pressure, especially after DeepSeek’s disruptive move and OpenAI’s recent free‑access announcements. Baidu’s CEO Robin Li, speaking at the World Governments Summit 2025 in Dubai, emphasized that innovation cannot be planned, noting that LLM inference costs have been dropping by over 90 % per year—far faster than the historical Moore’s law pace—and that cost reduction drives productivity and rapid innovation in Chinese enterprises.

Looking ahead, Baidu plans to launch several new models this year and expects a major multimodal upgrade in the upcoming Ernie 5.0 version.