Baidu Releases Wenxin 5.1: Search Capability Tops Domestic Charts, Pre-training Cost Only 6% of Industry Average
Baidu has officially launched its new generation foundational large model, Wenxin 5.1. Using 'multi-dimensional elastic pre-training' technology, it achieves leading basic performance with only about 6% of the pre-training cost of comparable models. It has topped the LMArena search ranking in China and ranks fourth globally. Agent capabilities have significantly improved, surpassing DeepSeek-V4-Pro, and creative writing is on par with Gemini 3.1 Pro.
Article intelligence
Key points
- Wenxin 5.1 uses multi-dimensional elastic pre-training, reducing pre-training cost to 6% of industry average for similar scale models.
- Topped LMArena search ranking in China with a score of 1223, fourth globally, the only domestic model on the leaderboard.
- Agent capabilities surpass DeepSeek-V4-Pro; creative writing matches Gemini 3.1 Pro.
- Now available on Baidu Qianfan model plaza and Wenxin Yiyan official website.
Why it matters
This matters because wenxin 5.1 uses multi-dimensional elastic pre-training, reducing pre-training cost to 6% of industry average for similar scale models.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
On May 9, 2026, Baidu officially released its next-generation foundational large model, Wenxin 5.1. This model employs a groundbreaking 'multi-dimensional elastic pre-training' technique, which allows it to achieve leading basic performance with only about 6% of the pre-training cost required by other models of similar scale. The innovation significantly reduces the financial barrier to high-performance AI development.
Wenxin 5.1 has demonstrated outstanding performance across multiple industry benchmark tests, excelling in areas such as agent capabilities, knowledge, reasoning, and deep search. Notably, its agent capabilities have shown substantial improvement, surpassing DeepSeek-V4-Pro. The model's creative writing ability is on par with Gemini 3.1 Pro, and its reasoning capabilities are approaching those of leading closed-source models.
In the latest LMArena rankings, Wenxin 5.1 scored 1223 points, taking first place among domestic models in the search category and ranking fourth globally. It is the only domestic model to appear on this leaderboard. The search capability refers to the model's ability to quickly retrieve, integrate, and generate information from multiple sources, producing coherent and reliable answers. This makes Wenxin 5.1 particularly suitable for complex business scenarios such as content generation, intelligent assistants, enterprise knowledge management, and agent applications.
The significant improvement in overall performance is attributed to the 'multi-dimensional elastic pre-training' technology, first introduced with Wenxin 5.0. This technique enables the training of multiple model scales in a single run. Wenxin 5.1 fully inherits the knowledge from Wenxin 5.0 while compressing total parameters to about one-third and active parameters to about half. As a result, it achieves leading basic performance with only about 6% of the pre-training cost of comparable models.
Wenxin 5.1 is now available on Baidu's Qianfan model plaza and the Wenxin Yiyan official website, offering open access to enterprise users and developers. The upcoming Create 2026 Baidu AI Developer Conference, scheduled for May 13-14 at the Beijing National Convention Center Phase II, will showcase the latest advancements in AI technology and industrial applications based on the Wenxin model.