Baidu Releases ERNIE-4.5-VL-28B-A3B-Thinking Multimodal AI Model

Baidu has introduced a new open-source multimodal AI model called ERNIE-4.5-VL-28B-A3B-Thinking, announced on its AI Studio platform. The model is designed to handle text, images, and video inputs while consuming significantly fewer computing resources than comparable systems from other major AI developers.

The model operates on a Mixture-of-Experts architecture with 28 billion total parameters but activates only 3 billion during inference. This selective activation allows it to perform complex reasoning tasks efficiently on a single 80GB GPU. Baidu states that the model matches the performance of leading systems while maintaining lower computational costs.

ERNIE-4.5-VL-28B-A3B-Thinking introduces several capabilities including visual reasoning, STEM problem solving, visual grounding, and video understanding. A distinctive feature, called “Thinking with Images,” enables the model to zoom in and out of images dynamically, improving its ability to analyze fine-grained visual details. It also supports tool calling for functions like image search and external data access.

The model is available under the Apache 2.0 license, allowing unrestricted commercial use. It supports deployment through multiple frameworks such as Transformers, vLLM, and Baidu’s FastDeploy toolkit. Developers can also fine-tune the model using ERNIEKit, Baidu’s training framework built on PaddlePaddle.

According to the official documentation, the model’s context window extends to 128,000 tokens and it supports both Chinese and English. Its design aims to make advanced multimodal reasoning more accessible to enterprises and researchers seeking efficient AI solutions.