Baidu released the Ernie 4.5 Series of Artificial Intelligence (AI) models in Open-SUS on Monday. The Chinese tech veteran had earlier said that it would make its ownership big language model (LLM) available to the open community on 31 July. It has now issued 10 different variants of the series, each model is built on mixture-off-exparts (MOE) architecture. Along with models, the company has also issued multi-hardeware development toolkit for Erni 4.5 in Open Source.
Baidu issues 10 variants of ERNIE 4.5 AI model in open source
One in Post On X (East was known as Twitter), the Chinese tech veteran announced the release of the 10 Open-SURS EERI 4.5AI model. Four of them are multimodal vision-language models, eight mo models, and two are thinking or reasoning models. Additionally, the list also includes five post-informed models, while others are pre-educated. These models can now be downloaded from the company’s hug face Entry Or from its github Entry,
One in blog postBaidu said that the Moe model has a total of 47 billion parameters, of which three billion are active at a time. The largest model in 10 variants has 424 billion parameters. All of them are trained using paddlepadal deep learning framework.
Based on the internal trial, the company claimed that the Erni-4.5-300B-A47B-BASE model to cross the Dipsek-V3-671 B-A37B-base at 22 out of 22 out of 28 benchmarks. Similarly, it was claimed that the Erny-4.5-21B-A3B-BASE outparforms Qwen3-30B-A3B-BASE despite 30 percent less parameters on many mathematics and logic benchmarks.
Baidu also revealed its training methods on model pages. The company enhances the model using an odd MOE structure in the pre-training process and using techniques such as intra-nod expert parallelism, memory-skilled pipeline scheduling, FP8 mixed-collective training, and a fine-dancing regeneration method.
Apart from models, Baidu has also made ERNIEKIT available for open community. It is a growth toolkit for Erni 4.5 series model. With this, developers can demonstrate pre-training, supervised fine-tuning (SFT), low-rank adaptation (Lora), and other adaptation techniques. In particular, all models permissible are available under Apache 2.0 license, which allows for both academic and commercial use.