ERNIE 4.5 VL 28B A3B
Baidu · 2025-06-30 · 29.4B parameters
ERNIE 4.5 VL 28B A3B is Baidu's post-trained multimodal Mixture-of-Experts chat model, based on ERNIE-4.5-VL-28B-A3B-Base and released here as PaddlePaddle weights. The checkpoint contains 29,399,294,336 parameters; Baidu labels it as 28B total and reports 3B activated per token. It accepts text, images, and video, combines a 28-layer vision-language backbone with a 32-layer vision encoder, supports switchable thinking and non-thinking modes over a 131,072-token context, and was post-trained through supervised fine-tuning and reinforcement learning.
Benchmark scores
| Benchmark | Score |
|---|---|
| SnakeBench Average Apples | 0.4 |
| SnakeBench Best Apples | 3.0 |
| SnakeBench Rating | 13.9 |
| SnakeBench Total Apples | 9.0 |
| SnakeBench Win Rate | 28.6 |