Wenxin 5.1 officially launched: parameters reduced to one-third of 5.0, pre-training costs only 6% of industry peers

According to Beating Monitoring, Baidu Wenxin Large Model 5.1 is officially launched, and users can experience it on the Wenxin Yiyan official website. Enterprises and developers can call the API through the Qianfan platform. Wenxin 5.1 is based on Wenxin 5.0 released in January of this year, with the core selling point being a significant reduction in model size and training costs: total parameters compressed to about one-third of 5.0, active parameters about half, and pre-training computational cost only 6% of models of the same scale.

The cost reduction comes from Baidu’s proposed Once-for-All elastic training framework. Traditional methods require pre-training for different model sizes separately. Wenxin 5.0 only needs to be pre-trained once, and can optimize a large number of different-sized sub-models 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 needed for training from scratch.

Previously, on April 30, the Wenxin 5.1 Preview version ranked first domestically with 1476 points on the LMArena text leaderboard. The official version further refreshed the benchmark: scoring 99.6 on the AIME26 math competition (using tools), second only to Gemini-3.1 Pro; the Agent capability surpasses DeepSeek-V4-Pro; Arena search leaderboard ranks fourth globally with 1223 points, first domestically.