Verified project record

PaddlePaddle/ERNIE

paddlepaddle/ernie provides a family of open-source large-scale multimodal models and an industrial-grade toolkit for resource-efficient training, compression, and deployment. It features a heterogeneous MoE architecture designed for high parameter efficiency across text and vision-language tasks.

Project overview

The project stands out for its heterogeneous modality structure that enables parameter sharing across modalities with modality-isolated routing, paired with a complete toolkit covering full-cycle development and multi-hardware inference.

Project type: Model Development · Model Runtime · Image & Vision
Deployment: Refer to project documentation
License: Apache-2.0

Best for

AI engineers and researchers who need to train, compress, and deploy large-scale multimodal models with high parameter efficiency.
Developers seeking an end-to-end toolkit for full-cycle multimodal model development and multi-hardware inference.

Key capabilities

A family of 10 large-scale multimodal models comprising MoE and dense architectures designed for text and vision-language tasks.
A Python API for executing offline model inference using FastDeploy.
Jupyter notebooks and Gradio demos showcasing best practices for conversations, web search, knowledge retrieval, and SFT/DPO tutorials.

Getting started

Setting up the project requires installing ERNIEKit, downloading model weights from the Hugging Face Hub, and executing training or inference via CLI or Python scripts.

Evidence and sources

GitHub project description: The official repository for ERNIE 4.5 and ERNIEKit – its industrial-grade development toolkit based on PaddlePaddle.
README: [2025-09] 🔥 Released ERNIEKit v1.2: - **New Features** - [WebUI] Added support for training and conversation functionalities with ERNIE 28b/424b VL models.
README: We introduce ERNIE 4.5, a new family of large-scale multimodal models comprising 10 distinct variants.
README: To achieve this without one modality hindering the learning of another, we designed a *heterogeneous MoE structure*, incorporated *modality-isolated routing*, and employed *router…
README: ERNIE-4.5-300B-A47B-Base surpasses DeepSeek-V3-671B-A37B-Base on 22 out of 28 benchmarks, demonstrating leading performance across all major capability categories.

Reviewed at: 2026-07-18T06:02:21Z