ERNIE 4.5 VL 28B A3B Base
ERNIE 4.5 VL 28B-A3B combines visual inputs with language in a sparse model. It produces text for image-based understanding tasks. This entry covers the base checkpoint.
Model design and access
This pretrained multimodal checkpoint precedes the chat and Thinking releases. It provides the vision-language expert architecture for further adaptation, with the visual and text components supplied together.ERNIE-4.5-VL-28B-A3B-Base-PT official model card
Practical use
Practical guidance. For a specialist visual domain, preserve the model’s image preprocessing during adaptation. Test on unseen layouts and capture conditions. A checkpoint that learns one document template can still fail when the same fields move to another part of a page.ERNIE-4.5-VL-28B-A3B-Base-PT official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ERNIE-4.5-VL-28B-A3B-Base-PT |
| Context window | 131,072 tokens Official checkpoint context window. Input and generation share this limit. ERNIE-4.5-VL-28B-A3B-Base-PT official model card |
| Parameters | 28B total parameters; 3B active per tokenERNIE-4.5-VL-28B-A3B-Base-PT official model card |
| Architecture | 28 layers; vision and language expert routingERNIE-4.5-VL-28B-A3B-Base-PT official model card |
| License | Apache 2.0ERNIE-4.5-VL-28B-A3B-Base-PT official model card |
Related models
- ERNIE 4.5 VL 28B A3B
Baidu · ERNIE 4.5
Post-trained sparse vision-language model with 28 billion total parameters and three billion active parameters.
Launch date not confirmedDownloadable