Ernie-4.5-VL-28B

Ernie-4.5-VL-28B-A3B-Thinking is Baidu's open vision-language model, released November 2025. It pairs a 21B-parameter MoE language backbone with a 7B vision encoder for 28B total parameters and only 3B active per token, which keeps it runnable on a single 80GB GPU. The Thinking variant does multi-step chain-of-thought over visual inputs, and a signature feature is thinking with images: the model can call tools to zoom into image regions and search for related images mid-reasoning. Baidu reports 87.1% on ChartQA, ahead of Gemini 2.5 Pro and GPT-5 at high reasoning, plus 858 on OCRBench and 82.5 on MathVista; these are team-reported. Context is 128K tokens across text, image, and video input. It ships under Apache 2.0 on Hugging Face with Transformers and vLLM support, making it one of the few high-performing VLMs in this class with no commercial licensing restriction. Training used multimodal RL with GSPO and IcePop to stabilize the MoE.

Team-reported: ChartQA 87.1, OCRBench 858, MathVista 82.5