On Monday, Baidu unveiled its Ernie 4.5 series of artificial intelligence (AI) models, releasing them as open-source. The announcement follows the company’s commitment to share its proprietary large language models (LLMs) with the open-source community, which was previously stated for July 31. The new series includes 10 distinct variants, all developed using the Mixture-of-Experts (MoE) architecture. In addition to these models, Baidu has also launched multi-hardware development toolkits specifically for Ernie 4.5 as open source.
Baidu Releases 10 Variants of Ernie 4.5 AI Models in Open Source
In a recent post on X (formerly Twitter), Baidu announced the release of the 10 open-source Ernie 4.5 AI models. Among these, four are multimodal vision-language models, eight utilize the MoE architecture, and two are designed for reasoning and cognitive tasks. Additionally, the release includes five post-trained models alongside several pre-trained variants. Interested users can download these models through Baidu’s Hugging Face listing or its GitHub repository.
Baidu elaborated on the capabilities of the MoE models in a blog post, noting that they consist of 47 billion parameters, with only three billion active at any single time. The most extensive model in the series boasts 424 billion parameters, and all variants have been trained utilizing the PaddlePaddle deep learning framework.
According to internal testing conducted by the company, the Ernie-4.5-300B-A47B-Base model outperforms the DeepSeek-V3-671B-A37B-Base in 22 of 28 benchmarks. Additionally, the Ernie-4.5-21B-A3B-Base demonstrated superior performance compared to Qwen3-30B-A3B-Base on various mathematics and reasoning tests, despite containing 30% fewer parameters.
Baidu has also disclosed details regarding its training methodologies on the model pages. The training process employed a heterogeneous MoE structure, integrating advanced techniques such as intra-node expert parallelism, memory-efficient pipeline scheduling, FP8 mixed-precision training, and a fine-grained recomputation strategy.
Alongside the models, Baidu has introduced ErnieKit, a development toolkit designed for the Ernie 4.5 series. This toolkit enables developers to carry out tasks such as pre-training, supervised fine-tuning (SFT), and Low-Rank Adaptation (LoRA), among other customization options. All models are distributed under the permissive Apache 2.0 license, facilitating use in both academic and commercial contexts.
