The brief
- ERNIE 5.0 (2.4 trillion parameters) was the only Chinese model in LMArena's text top 10 in January 2026.
- ERNIE 5.1 (May 2026) is about a third the size of 5.0; Baidu says pre-training cost about 6% of comparable models.
- The Ernie assistant had 202 million monthly users in December 2025; Baidu said ERNIE served 16.5 billion API calls a day in January 2026.
Technical approach
One model for text, images, sound and video
ERNIE 5.0 is trained on all four kinds of data together from the start, instead of bolting separate vision or audio models onto a text model.
A huge model that wakes only a sliver
ERNIE 5 is a mixture of experts: of its 2.4 trillion parameters, under 3% are used for any one answer, which keeps the cost of each reply down.
Many model sizes from one training run
ERNIE 5.1 was cut from ERNIE 5.0 with a method that trains many model sizes at once; Baidu claims 6% of comparable pre-training cost.
Search as the delivery channel
ERNIE answers inside Baidu Search and the Ernie assistant, which can book flights or order food through apps like JD.com, Meituan and Trip.com.

