Baidu releases Ernie 5.1, a smaller language model with pre-training costs at six percent of comparable models, scoring 1,223 points on Arena Search Leaderboard

Baidu released Ernie 5.1, built on Ernie 5.0 with roughly a third of the total parameters and about half the active parameters per query. Pre-training costs came in at six percent of what comparable models require, according to Baidu.

Ernie 5.1 uses a 'Once-For-All elastic training framework' that optimizes an entire family of differently sized models in a single training run, avoiding separate expensive pre-training passes for each model size. Baidu also rebuilt its reinforcement learning infrastructure with separate subsystems (model updates, response generation, and evaluation) that scale independently.

On the Arena Search Leaderboard as of May 9, Ernie 5.1 scored 1,223 points in 4th place globally and 1st among Chinese models. On autonomous AI agent tasks, Baidu claims Ernie 5.1 beats DeepSeek-V4-Pro; on knowledge and reasoning benchmarks (GPQA, MMLU-Pro), it comes close to Google's Gemini 3.1 Pro.

Ernie 5.1 is available through ernie.baidu.com and Baidu AI Studio playground, with rollout to more than ten creative platforms including Isekai Zero, Mulan AI, Diting Huanliu, and Storymaster. Model weights have not been released.