The model combines an InternViT-300M vision encoder with Qwen2-0.5B-Instruct and has an 8K context window. It targets tasks including OCR, document and chart analysis, and video question answering.

According to the source, images can be split into up to 12 tiles at 448×448 pixels. OpenGVLab says the model matches or exceeds some 2B–4B models on selected visual tasks; these are task-specific reported results.

Why it matters

Its compact size and reported OCRBench score make it relevant to memory-constrained, local vision-language applications. The results are not presented as independently verified.

GUIDES

Something to try

Hypothetically, compare OCR and chart-interpretation results on your own documents against a larger vision-language model.