Baidu open-sources the ERNIE 4.5 model family
The ten variants include MoE models with 47B and 3B active parameters (up to 424B total) and a 0.3B dense model, reversing Baidu's prior closed-weight strategy for its flagship line.
- Open weights & ecosystem
- Models & capabilities
- Notable
Baidu open-sourced its ERNIE 4.5 model family, releasing ten variants under the Apache 2.0 licence on Hugging Face, GitHub and its own PaddlePaddle platform. The family ranges from a 0.3-billion-parameter dense model up to mixture-of-experts models with 47 billion and 3 billion active parameters and as many as 424 billion total parameters, and includes multimodal variants alongside text-only ones.
The release marked a reversal for Baidu, which had until then kept its flagship ERNIE models closed, accessible only through its own API and Wenxin chatbot product, in contrast with China’s other major labs. That strategy had come under pressure over the preceding months as DeepSeek’s open-weight R1 model drew global attention and adoption largely because it was freely downloadable, prompting several Chinese labs, including Alibaba’s Qwen team, to lean further into open releases as a way of building developer mindshare that a closed API could not.
Baidu’s own materials claimed strong benchmark performance for the larger variants relative to other open models available at the time, though independent verification varied release to release across the crowded mid-2025 field of open Chinese models. The release added ERNIE to a run of open-weight launches that month — MiniMax’s M1 two weeks earlier and, within days, further Chinese-lab releases — that collectively made open weights, rather than closed APIs, the default way Chinese labs competed for international developer attention, even as the underlying training compute for the largest variants remained subject to US export controls on advanced chips.