Baidu has released a new open source multimodal model called ERNIE 4.5 VL 28B A3B Thinking. The company says the model can match or beat much larger rivals on many vision and document tasks while using far less compute.
Model and design
ERNIE 4.5 VL 28B A3B Thinking is part of Baidu’s ERNIE 4.5 family. The model uses a mixture of experts design. Only about three billion parameters are active for each input while the full model holds many more parameters in reserve. This routing approach cuts runtime cost while keeping large model capacity in the background.
Baidu calls one feature Thinking with Images. The model can zoom into image regions and then zoom back out. This lets the system combine broad scene context with fine visual detail. The model also facilitates visual grounding and tool requests like image search to make it be able to retrieve additional information where necessary.
Baidu published the model as under Apache 2.0 license. The model files and sample code are available on Hugging Face and in Baidu toolkits. The company also provides support for common runtimes such as Transformers vLLM and FastDeploy. Baidu states that a single 80 gigabyte GPU is enough to run inference with the model in many settings.
The paper and technical notes explain the family design and show training strategies that stabilize mixture of experts training. The technical report also documents options for compression and integration with ERNIEKit. These resources help teams that want to test the system on their own data.
Performance and use
Baidu shared benchmark slides that show strong results on document understanding chart analysis and visual reasoning. The company compared the model to Google and OpenAI systems and reported higher scores on several test sets. Independent reviews and full public audits are still limited. Readers should treat the vendor charts as an important data point that needs wider validation.
The model’s strengths appear to be tasks that need tight alignment of image detail and structured text. Examples include reading complex charts extracting tables from images and solving diagrams in STEM problems. Baidu also highlights temporal awareness for short videos and tool enabled flows that call image search or zoom functions. These features suit enterprise work such as document processing quality control and image based customer support.
There are practical limits to consider. The model still needs a high memory GPU for full size inference. Video workloads add cost and complexity. The mixture of experts routing adds deployment complexity compared with simpler dense models. Safety testing bias audits and robust failure mode analysis are not fully detailed in the public materials. Teams should run their own tests on real production data before trusting the model for high risk work.
The open license and the single card claim lower the bar for trials. That can make the model attractive for mid market firms and research labs that need strong vision language ability but lack massive GPU farms. At the same time enterprise adopters must weigh integration work monitoring and governance needs that come with any cutting edge model deployment.