Apple has released LensVLM-9B, a 9-billion-parameter Vision Language Model designed to process compressed images of text more efficiently. The model scans these compressed inputs and uses learned tools to selectively expand only the relevant pages to their uncompressed form.
The release includes the model weights under the Apple Machine Learning Research Model License and source code under the Apple Sample Code License. It is built upon modifications to the Qwen architecture and is accompanied by a paper titled "LensVLM: Selective Context Expansion for Compressed Visual Representation of Text" available on arXiv.
This approach allows the model to handle visual text representations by focusing computational resources only on necessary details, potentially improving efficiency in processing large documents.