ERNIE-5.1-Preview ranks 13th on LMArena Text
Baidu's ERNIE-5.1-Preview placed 13th overall and first among Chinese models on LMArena's text leaderboard, which suggests the smaller, cheaper-to-train model can compete with frontier systems in head-to-head user votes.
Baidu's ERNIE-5.1-Preview ranked No. 13 globally and No. 1 among Chinese models in the LMArena Text leaderboard update released on April 30, Baidu said in a research post.
Baidu also listed the model's results on several LMArena category leaderboards:
- Legal & Government: #1 globally
- Business, Management & Financial Ops: #4 globally
- Software & IT Services: #7 globally
- Math: #9 globally
As we reported earlier, ERNIE 5.1 is a smaller and cheaper successor to ERNIE-5.0. Baidu says the Preview model builds on ERNIE-5.0's pre-training foundation but has about one-third of the total parameters and about half of the active parameters. According to the company, it reaches leading performance for its size with roughly 6% of the pre-training cost of comparable models. Baidu credits the gains to "decoupled fully-asynchronous reinforcement learning" and scaled agentic post-training.
The post does not give LMArena Elo scores, vote counts, confidence intervals, or the models ranked above ERNIE-5.1-Preview. It also does not name the comparable models used for the cost comparison or give absolute parameter counts. LMArena rankings come from crowdsourced pairwise preference votes, and placements can change as new models are added and more votes come in.
Baidu says the ERNIE models are available to try at ernie.baidu.com.