Baidu released the Ernie 4.5 series of artificial intelligence (AI) models in open-source on Monday. The Chinese tech giant had previously said that it would make its proprietary large language models (LLMs) available to the open community on July 31. It has now released 10 different variants of the series, with each of the models built on Mixture-of-Experts (MoE) architecture. Alongside the models, the company has also released multi-hardware development toolkits for Ernie 4.5 in open source.
Baidu Releases 10 Variants of Ernie 4.5 AI Models in Open Source
In a post on X (formerly known as Twitter), the Chinese tech giant announced the release of the 10 open-source Ernie 4.5 AI models. Four of them are multimodal vision-language models, eight are MoE models, and two are thinking or reasoning models. Additionally, the list also includes five post-trained models, while others are pre-trained. These models can now be downloaded from either the company’s Hugging Face listing or from its GitHub listing.
In a blog post, Baidu said that the MoE models feature a total of 47 billion parameters, with three billion of them being active at a time. The largest models among the 10 variant features 424 billion parameters. All of them are trained using the PaddlePaddle deep learning framework.
Based on internal testing, the company claimed the Ernie-4.5-300B-A47B-Base model surpasses DeepSeek-V3-671B-A37B-Base on 22 out of 28 benchmarks. Similarly, it claimed that the Ernie-4.5-21B-A3B-Base outperforms Qwen3-30B-A3B-Base on multiple mathematics and reasoning benchmarks despite having 30 percent fewer parameters.
Baidu also revealed its training methods on the model pages. The company used a heterogeneous MoE structure in the pre-training process and scaled the models using techniques such as intra-node expert parallelism, memory-efficient pipeline scheduling, FP8 mixed-precision training, and a fine-grained recomputation method.
Apart from the models, Baidu has also made ErnieKit available to the open community. It is a development toolkit for the Ernie 4.5 series models. With this, developers can perform pre-training, supervised fine-tuning (SFT), Low-Rank Adaptation (LoRA), and other customisation techniques. Notably, all the models are available under the permissive Apache 2.0 licence, which allows for both academic and commercial usage.
