Baidu ERNIE 5.1 Review: 94% Compute Reduction, #4 Search Arena, AIME26 at 99.6

Baidu ERNIE 5.1 (May 8, 2026) is a closed MoE model derived from ERNIE 5.0's 2.4T-parameter architecture using a novel 'Once-For-All' elastic training framework. Total parameters compress to ~800B (~1/3 of 5.0), active parameters to ~half — at only 6% of the pre-training compute of comparable models. It ranks #4 globally on LMArena Search Arena (#1 Chinese model) and #2 on AIME26 math (99.6). But MMLU-Pro comes in last among frontier comparisons, coding is weak, no weights are released, and Baidu API access requires a Baidu account. Rating: 3.5/5.