ERNIE 4.5 VL 28B A3B Thinking
The visual Thinking release extends the 28B-A3B architecture with visual reasoning data and reinforcement learning. It supports visual grounding, video understanding and image tools.
Model design and access
The model card describes GSPO and IcePop training techniques with dynamic task difficulty. Image zoom and image search are external tools that can support its reasoning. Applications must supply those tools; the weights do not perform independent network access.ERNIE-4.5-VL-28B-A3B-Thinking official model card
Practical use
Practical guidance. Use it when an answer needs to locate and inspect a small visual feature. Retain tool inputs and outputs so that you can check whether a wrong answer came from image selection, zooming or reasoning. Evaluate video questions with accurate timestamps.ERNIE-4.5-VL-28B-A3B-Thinking official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ERNIE-4.5-VL-28B-A3B-Thinking |
| Context window | 131,072 tokens Official checkpoint context window. Input and generation share this limit. ERNIE-4.5-VL-28B-A3B-Thinking official model card |
| License | Apache 2.0ERNIE-4.5-VL-28B-A3B-Thinking official model card |
| Documented deployment example | At least 48GB GPU memory for the cited single-card setupERNIE-4.5-VL-28B-A3B-Thinking official model card |
Related models
- ERNIE 4.5 VL 28B A3B
Baidu ยท ERNIE 4.5
Post-trained sparse vision-language model with 28 billion total parameters and three billion active parameters.
Launch date not confirmedDownloadable