Baidu Releases ERNIE 5.1 with Lower Training Cost

Per Baidu's official blog post dated May 9, 2026, Baidu released ERNIE 5.1 as a successor built on the pretraining foundation of ERNIE 5.0, claiming total-parameter compression to approximately one-third and active-parameter reduction to about one-half. Baidu's announcement states the model achieves flagship-level performance using about 6% of the pretraining cost of comparable models and reports a score of 1,223, ranking ERNIE 5.1 4th globally and 1st among Chinese models on the Arena Search leaderboard (Per Baidu's blog). Decrypt summarizes the cost claim as a 94% reduction relative to rival training costs. I-Scoop's coverage and Baidu's blog describe the model as selected from an elastic Once-For-All-style training family derived from ERNIE 5.0 (I-Scoop; Baidu blog).

Per Baidu's technical writeup, ERNIE 5.1 was produced by extracting an optimized subnetwork from a larger elastic pretraining run rather than by training an independent model from scratch. Baidu reports a new disaggregated fully-asynchronous reinforcement learning infrastructure intended to address training-inference divergence, resource utilization, and long-tail effects. The blog also attributes additional capability gains to a scaled agentic post-training stage and an end-to-end synergy strategy across environment, expert, and integration phases (Baidu blog).

Industry-pattern observations: Reusing a larger pretrained family and selecting performant subnetworks is consistent with recently explored efficiency strategies such as Once-For-All networks, mixture-of-experts pruning, and sparse activation. These approaches aim to trade raw parameter count for effective active capacity, which can reduce compute during pretraining and inference while preserving task performance. If the method generalizes beyond benchmark tasks, practitioners could expect lower-cost paths to deployable foundation models for many production workloads.

Editorial analysis: The headline claim, achieving leaderboard-grade performance at a fraction of pretraining cost, challenges the dominant scaling narrative that equates progress strictly with ever-larger monolithic pretraining runs. For ML engineers and infrastructure teams, a validated ~94% training-cost reduction would materially change cost projections for model development and might shift investments toward flexible training pipelines, elastic model families, and more aggressive subnetwork selection or post-training tuning.

Editorial analysis: Observers should look for third-party benchmark reproductions, independent inference-cost comparisons on real-world workloads, and open-source evaluation artifacts. Also watch for technical writeups or code releases that clarify the subnetwork selection criteria, the specifics of the disaggregated asynchronous RL pipeline, and any trade-offs in robustness or long-tail task handling reported by external evaluators.