ERNIE 5.1 is here with a third of the params and a fraction of

Baidu's ERNIE 5.1 is now live. You can try it on the ERNIE Bot website, and developers can hit the API through the Qianfan platform. The model is trained from ERNIE 5.0, which launched back in January, and its whole selling point is dramatic compression: total param count is slashed to about one-third of 5.0, activated parameters are roughly half, and pre-training compute cost is just 6% of what similarly sized models demand.

That cost reduction comes from Baidu's Once-for-All elastic training framework. Normally you'd run separate pre-training runs for models of different sizes. ERNIE 5.0 does one pre-training pass, then dynamically samples and optimizes a whole bunch of sub-models simultaneously. ERNIE 5.1 is the optimal structure extracted from that sub-model matrix — it inherits 5.0's knowledge directly, skipping the need to train from scratch.

Back on April 30th, the ERNIE 5.1 Preview hit 1476 on LMArena's text leaderboard, ranking first in China. The official version pushes those numbers higher: on the AIME26 math competition (with tools) it scored 99.6, second only to Gemini 3.1 Pro. Its agent capabilities beat DeepSeek-V4-Pro. On Arena's search leaderboard, it's fourth globally and first in China with 1223 points.

These results put ERNIE 5.1 ahead of many rivals, including Zhipu AI's GLM-5V-Turbo in certain coding benchmarks.