技术规格

ERNIE-4.5-VL-28B-A3B-Base

ERNIE-4.5-VL-28B-A3B-Base is a multimodal Mixture-of-Experts (MoE) foundation model developed by Baidu as part of the ERNIE 4.5 model family. Specifically engineered for sophisticated vision-language tasks, the model integrates 28 billion total parameters while activating only 3 billion parameters per token during inference. This sparse activation strategy allows the model to maintain the extensive knowledge capacity of a larger system while significantly reducing the computational overhead and latency typically associated with high-parameter models. It is designed to process and synthesize information across multiple modalities, including text, images, and video, supporting a substantial context length of up to 131,072 tokens.

The technical architecture of the ERNIE-4.5-VL series introduces a heterogeneous MoE structure that facilitates both parameter sharing across modalities and the use of dedicated parameters for individual modalities. Key innovations include modality-isolated routing, which prevents interference between textual and visual learning, as well as router orthogonal loss and multimodal token-balanced loss mechanisms to ensure stable expert utilization. The model employs Grouped-Query Attention (GQA) for efficient memory management and utilizes Rotary Position Embeddings (RoPE) to handle extended context windows. Training is conducted within the PaddlePaddle deep learning framework using advanced parallelization strategies, including intra-node expert parallelism and FP8 mixed-precision training.

In operation, the ERNIE-4.5-VL-28B-A3B-Base serves as a versatile backbone for applications requiring high-fidelity cross-modal reasoning. It supports distinct functional modes, including a "thinking" mode for enhanced logical reasoning and a "non-thinking" mode optimized for perceptual tasks such as document analysis, optical character recognition (OCR), and visual knowledge retrieval. Its capabilities extend to agentic interactions, where it can utilize external tools for fine-grained image zooming or search. The model is released with open weights under the Apache 2.0 license, providing a flexible resource for developers and researchers to deploy multimodal solutions across various hardware platforms.

没有可用的 ERNIE-4.5-VL-28B-A3B-Base 评估基准。