Apple publishes model that searches compressed document pages

Apple has made files and code available for LensVLM, a 9B vision-language research model that scans compressed images of text and expands pages it identifies as relevant. The model is intended to answer questions about long documents without reading every page at full resolution. Its published demo accepts a text file and a question, with compression settings of 5x, 10x or 15x. Apple provides the model files under its Machine Learning Research Model License and the accompanying source code under a separate Sample Code License. The release supplies a runnable research model, though the model card does not establish performance in production use.