AsianFin — Baidu has officially open-sourced its Wenxin Large Language Model 4.5 series, releasing a total of 10 models including Mixture-of-Experts (MoE) models with 47B and 3B active parameters, and a dense model with 0.3B parameters. The release includes both pretrained weights and inference code.

The open-source Wenxin 4.5 models are now available for download and deployment on platforms such as PaddlePaddle’s Xinghe Community and Hugging Face. In addition, API access to the models is now available through Baidu’s Qianfan Model-as-a-Service platform on Baidu AI Cloud.

Baidu first announced its intention to launch and open-source the Wenxin 4.5 models in February this year, confirming today’s official release date of June 30.