Baidu Releases ERNIE 5.1 with Pre-Training Cost at 6% of Industry Peers

Baidu officially released its ERNIE 5.1 foundation model on May 9, 2026, achieving a breakthrough in training cost efficiency. The model's pre-training cost is approximately 6% of comparable models at similar scale, using a multi-dimensional elastic pre-training technique that compresses total parameters to roughly one-third and active parameters to one-half versus ERNIE 5.0. On the LMArena search leaderboard, ERNIE 5.1 scored 1,223 points, ranking fourth globally and first among Chinese models. Its agent capabilities surpass DeepSeek-V4-Pro, and its creative writing and reasoning approach Gemini 3.1 Pro. In AIME26 math competition scoring, it achieved 99.6, second only to Gemini 3.1 Pro. The model is available on Baidu's Qianfan platform and ERNIE website for enterprises and developers. Baidu's AI Developer Conference is scheduled for May 13.