Baidu Launches ERNIE 5.1 Using Sub-Network Extraction Architecture

Baidu released ERNIE 5.1 by extracting an optimal sub-network from its 2.4 trillion parameter ERNIE 5.0 model rather than training a new system from scratch. [1][2] This technique reduces total parameters to approximately one third of ERNIE 5.0 and active parameters to roughly one half. [1][2] The model utilized a "Once-For-All" elastic training framework, bringing pre-training compute costs down to approximately 6% of comparable models. [1][2] Additionally, Baidu built a decoupled fully asynchronous reinforcement learning infrastructure that separates training, inference, reward, and agent loop control systems. [2]

ERNIE 5.1 scored 1223 on the LMArena Search Arena, ranking fourth globally and first among domestic models. [1][2] During the AIME26 math competition evaluation using tools, the model scored 99.6, placing it behind Gemini 3.1 Pro. [2] For agent evaluation tasks including τ³-bench and SpreadsheetBench-Verified, ERNIE 5.1 outperformed DeepSeek-V4-Pro. [2] Ahead of the release announcement, Baidu's stock price rose by nearly 6% at the close on May 8. [2]