ERNIE 4.5

ERNIE 4.5 marked Baidu's shift to genuinely open releases: in June 2025 the company published a family of ten models under Apache-2.0, from a 0.3B dense model up to mixture-of-experts language models at 21B (3B active) and 300B (47B active) and vision-language models at 28B and 424B total parameters. The architecture uses a heterogeneous MoE design that shares some parameters across text and vision modalities while keeping others modality-specific, and supports text, image, and video input, a 128K context window, and thinking and non-thinking modes on the VLMs. The repository doubles as tooling: ERNIEKit handles SFT, LoRA, DPO, and quantization-aware training on PaddlePaddle, while FastDeploy covers serving with W4A16, W8A16, FP8, and even 2-bit formats. Base and post-trained checkpoints are all downloadable, and the Apache-2.0 terms permit commercial use. With around 7.7k GitHub stars and maintenance continuing into 2026, it is the main entry point into Baidu's open model ecosystem, though the PaddlePaddle-first stack is a real consideration for PyTorch-centric teams.