ERNIE 4.5 VL 424B A47B Base
The largest ERNIE 4.5 vision-language release provides a large sparse model for image and text understanding. This entry covers the base checkpoint.
Model design and access
The base checkpoint supplies the pretrained multimodal network before chat alignment. Both text and visual representations are part of the release. It is a large model-development resource, not an instruction-ready replacement for the chat checkpoint.ERNIE-4.5-VL-424B-A47B-Base-PT official model card
Practical use
Practical guidance. Use the base for controlled multimodal training when the larger expert capacity serves a measured need. Keep a general image-and-text evaluation set alongside the target domain so that specialized training does not hide losses in broader understanding.ERNIE-4.5-VL-424B-A47B-Base-PT official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ERNIE-4.5-VL-424B-A47B-Base-PT |
| Context window | 131,072 tokens Official checkpoint context window. Input and generation share this limit. ERNIE-4.5-VL-424B-A47B-Base-PT official model card |
| Parameters | 424B total parameters; 47B active per tokenERNIE-4.5-VL-424B-A47B-Base-PT official model card |
| Architecture | 54 layers; separate text and vision expert poolsERNIE-4.5-VL-424B-A47B-Base-PT official model card |
| License | Apache 2.0ERNIE-4.5-VL-424B-A47B-Base-PT official model card |
Related models
- ERNIE 4.5 VL 424B A47B
Baidu ยท ERNIE 4.5
Post-trained vision-language model with 424 billion total parameters and 47 billion active parameters.
Launch date not confirmedDownloadable