Baidu ERNIE 5.1 ranks 4th on Arena, cuts training costs

Baidu launched ERNIE 5.1 on May 9. The model scored 1,223 on the Arena leaderboard, placing fourth globally and first among Chinese models. Baidu reported it reached that position while spending about 6% of the pre-training budget that peer models typically use.

Baidu described the change as a gain in parameter efficiency. ERNIE 5.1 reduced its total parameter count by roughly two-thirds versus ERNIE 5.0 and activates about half of the remaining parameters during inference. The company said those choices cut compute needs in pre-training and lowered ongoing operational expense during inference. Baidu noted independent evaluations validated ERNIE 5.1 as achieving flagship-level performance but did not provide details of the tests.

The launch comes amid tighter U.S. export controls on advanced chips that have limited access to high-end GPUs for Chinese firms. Companies in China have faced pressure to improve model performance on constrained hardware, and Baidu framed ERNIE 5.1’s efficiency as a response to those supply limitations.

Baidu has not issued a token tied directly to ERNIE 5.1. Its blockchain initiative, Ernie Chain, went live in March 2026 and integrates ERNIE models to deliver on-chain AI services. Ernie Chain raised $50 million earmarked for decentralized finance applications. Some market analysts have forecast roughly a 25% price increase for efficiency-focused tokens such as TAO and point to ERNIE 5.1’s efficiency as one factor behind that outlook.

The Arena ranking places Baidu among the top developers on a widely used benchmark for large language models. The company emphasized reduced parameter activation during inference and a smaller total parameter count as the mechanisms for lower costs. Baidu did not disclose the full validation methods or the exact hardware and budget figures used in its comparison.