"According to Baidu, the model uses 28 billion total parameters, but only 3 billion are active during each use. That setup helps lower computing costs while keeping performance high. Baidu also notes that the model runs 2 to 3 times faster than others with a similar structure and needs less memory. It is available under the Apache 2.0 license, which allows companies and developers to use it for both personal and commercial projects."
"The ERNIE model targets specific tasks where it can handle both images and written content at the same time. It can zoom in on parts of an image, analyze details, and use those insights to form full answers. Baidu researchers say they used a large amount of training data that focused on visual tasks. They also applied reinforcement learning techniques to improve how the model solves harder problems. The company claims that the model performs well in visual reasoning, science, and math tasks."
Baidu introduced ERNIE-4.5-VL-28B-A3B-Thinking, an open-source multimodal AI model under the Apache 2.0 license. The model uses 28 billion parameters with 3 billion active per inference to reduce compute and memory costs. Baidu reports 2–3× faster runtime than comparable architectures. The release targets visual reasoning tasks combining images and text, enabling zoomed-in image analysis and detailed answers. Training emphasized visual-task data and reinforcement learning to improve complex problem solving. The model claims strong performance on charts, documents, image-based questions, science and math tasks, and video understanding. The launch heightens competition with closed U.S. AI offerings.
Read at TipRanks Financial
Unable to calculate read time
Collection
[
|
...
]