Intel / article

ERNIE-4.5-VL-28B-A3B-Thinking: A Breakthrough in Multimodal AI

Source: unknown
Date: Published Nov 11, 2025

Context

ERNIE-4.5-VL-28B-A3B-Thinking is designed to investigate visual details with tools as part of answering a question. Baidu's release article describes pairing the model with image zooming and image search, so a visual reasoning workflow can inspect a region more closely or seek information about unfamiliar content.

The upgrade combines mid-training on visual-language reasoning data with reinforcement learning on tasks whose answers can be checked. The article shows the model deciding on its own to call a zoom tool to inspect a sign and a search tool to identify a toy. It does not say whether those tools are built in or must be supplied by the application, so the capability should be read as what the model does when paired with such tools.

Terms in this piece · Glossary

multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.

Source link unavailable.unknown