文心 5.1 Officially Launched: Parameter Cut to One-Third of 5.0, Pre-training Cost Reduced to Only 6%

According to Dongzuo Beating monitoring, Baidu Wenxin Big Model 5.1 has officially launched. Users can experience it on the Wenxin Yanyan official website, and enterprises and developers can use the API through the Qianfan platform. Wenxin 5.1 is based on the Wenxin 5.0 released in January this year. The core selling point is a significant reduction in model size and training costs: the total parameter volume has been compressed to about one-third of 5.0, with activated parameters around half, and the pre-training computational cost is only 6% of a model of the same scale.

The cost reduction comes from Baidu's proposed Once-for-All elastic training framework. The traditional approach requires separate pre-training runs for models of different scales. Wenxin 5.0 only undergoes pre-training once, optimizing a large number of sub-models of different sizes simultaneously through dynamic sampling. Wenxin 5.1 is the optimal structure extracted from this sub-model matrix, directly inheriting the knowledge from 5.0, saving the computational power required for training from scratch.

On April 30th, the Wenxin 5.1 Preview version had already achieved the top ranking in the LMArena Text Ranking with a score of 1476, leading domestically. The official version further excelled in benchmark tests: scoring 99.6 in the AIME26 Mathematical Contest (Tool Assisted), second only to Gemini-3.1 Pro; surpassing Agent's capabilities in DeepSeek-V4-Pro; ranking fourth globally and first domestically in the Arena Search Ranking with a score of 1223.