Baidu officially released its new generation language model Ernie5.1 on May 11, 2026. This model is derived from the Ernie5.0 pre-training foundation, which was launched in January of this year and has 2.4 trillion parameters. Through an innovative "one-time elastic training framework," Baidu achieved the ability to optimize multiple model sizes in a single training run, reducing the pre-training cost of Ernie5.1 to only 6% of that of similar models.

QQ20260512-093146.jpg

As of May 9, Ernie5.1 ranked fourth globally and first among Chinese models on the Arena Search ranking list with 1223 points, demonstrating high resource utilization and performance balance capabilities.

At the architectural level, Ernie5.1 adopts a sub-model configuration with adjustable depth, width, and number of active experts. Its parameter count is only one-third of its predecessor, and the effective parameter count per query has also been reduced by about half. To overcome the "see-saw effect" in multi-skill training, Baidu applied a four-phase post-training process, using parallel specialized training code, inference, and proxy expert models, combined with strategy distillation and reinforcement learning, successfully solving the industry challenge of interference between programming ability and creativity. In addition, the reconstructed reinforcement learning infrastructure decouples model updates, response generation, and evaluation, and when combined with a standardized low-precision computing library, significantly improves the stability of large-scale training.

QQ20260512-093200.jpg