- Baidu has launched the Ernie 5.1 model, achieving a leapfrog optimization in parameter size and computing overhead.
- The model's search capabilities have topped domestic rankings, while its agent and reasoning performance rival top global competitors.
Baidu has officially released its Ernie 5.1 AI (artificial intelligence) model, slashing its pre-training cost to just 6% of industry models of similar scale and achieving significant optimization in parameter size and computing overhead.
While drastically reducing computing costs, the new model has successfully topped China's search capability rankings, according to an announcement on Saturday.
As Baidu's latest achievement in the artificial intelligence (AI) sector, Ernie 5.1 marks a major breakthrough in the cost-efficiency of large models.
Relying on innovative multi-dimensional elastic pre-training technology, Ernie 5.1 extracts the optimal sub-structures from the previous generation 5.0 model released last November.
Its total parameter count has been successfully compressed to one-third of the original, and the activation parameter count has been reduced to about one-half.
This technical approach allows Baidu to optimize multiple sub-models in a single training session, significantly lowering pre-training costs, the announcement said.
Despite the sharp drop in training costs, Ernie 5.1 continues to perform strongly across various authoritative benchmark tests.
Its agent capabilities have surpassed DeepSeek-V4-Pro, while its creative writing and reasoning levels are approaching top-tier international closed-source models such as Gemini 3.1 Pro.
In the LMArena evaluation, the model ranked fourth globally on the search leaderboard with a score of 1223, making it the only Chinese model to make the list.
To drive the evolution of large models toward autonomous decision-making agents, Baidu has built a decoupled, fully asynchronous reinforcement learning technology.
This architecture resolves global optimization challenges caused by low resource utilization and long-tail effects. Through a multi-stage reinforcement learning training pipeline, Ernie 5.1 ensures stability and outstanding performance when handling complex tasks.
Currently, Ernie 5.1 has been simultaneously launched on Baidu's Qianfan model platform and its official website, open to enterprises and developers, according to the announcement.
The market is closely watching the upcoming Baidu AI Developer Conference scheduled for May 13. At the event, Baidu founder Robin Li is expected to disclose more technical details and future commercialization plans for the model.
