Baidu and ERNIE: China's early mover in large language models

From search engine to generative AI

Baidu was founded in 2000 in Beijing’s Zhongguancun tech district. It started as a search engine and now describes itself as a full-stack AI company, from chips to applications. It was among the first Chinese tech giants to launch a consumer chatbot, powered by its ERNIE family of large language models (Wenxin in Chinese).

The ERNIE families

The lineup has grown with each generation:

• ERNIE 4.5: a series of language and vision models (ERNIE-4.5-VL) in a wide range of sizes, some built on a mixture-of-experts design;
• ERNIE X1: reasoning models that think before answering;
• ERNIE 5.0, then 5.1: the flagship models, presented as natively multimodal, handling text, images, audio and video in a single model.

Baidu also releases specialized models, for example for document recognition (PaddleOCR-VL) and image generation.

A mixed policy: open on one side, closed on the other

Baidu’s strategy has two layers. The ERNIE 4.5 generation is released as open weights under the Apache 2.0 license, which allows commercial use. It is available on Hugging Face (the baidu organization), on GitHub and on Baidu’s AI Studio platform.

The most advanced models, ERNIE 5.0 and 5.1, remain proprietary: you can only use them through Baidu’s services. In practice, if you want to host ERNIE on your own servers, the 4.5 generation is the one to look at.

How to access the models

There are three entry points. The ERNIE assistant at ernie.baidu.com lets you try the models without installing anything. Baidu AI Cloud’s Qianfan platform sells API access to businesses, with tools for building agents and knowledge bases. Finally, the open ERNIE 4.5 weights can be deployed with Baidu’s own ERNIEKit and FastDeploy, or with standard tools.

To compare ERNIE with other Chinese models, use our model comparison tool or read the profile of DeepSeek, whose models are also offered on Qianfan.