Skip to main content
Back

Apple's LensVLM-9B Reads Long Documents at 4.3x Compression

Apple's new 9B vision-language model reads compressed page thumbnails first, then zooms into only the pages it needs to answer your question.

  • Apple released LensVLM-9B, a Qwen3.5-9B finetune for compressed visual document understanding.
  • Renders text pages as thumbnails, then calls a read_page tool to expand only relevant pages.
  • Matches full-text accuracy at 4.3x compression, beats retrieval baselines up to 10.1x.
  • Supports configurable 5x, 10x, and 15x compression ratios at inference time.
  • Available under Apple ML Research license; works with Transformers, vLLM, and SGLang.
  • Full paper and code published by Apple's ML research team.

Apple has released LensVLM-9B, an open-weights vision-language model designed for question answering across long documents. Built on Qwen3.5-9B-Base, the 9-billion-parameter model scans compressed page images, identifies likely sources, and invokes tools to retrieve selected pages at full resolution.

Long PDFs consume substantial context when every page is tokenized as text. LensVLM reduces that cost by using page images as a compact visual index, then spending additional tokens only on material relevant to the question. Apple reports accuracy comparable to a full-text baseline at 4.3x effective compression across seven text question-answering benchmarks.

Detail LensVLM-9B
Base model Qwen3.5-9B-Base
Parameters 9 billion
Primary task Long-document question answering
Compression settings 5x, 10x, and 15x
Weight format BF16, approximately 18.8 GB
License Apple Machine Learning Research License

Pages become a visual index

LensVLM renders document text into images because a vision encoder converts each fixed-size input into a bounded sequence of visual tokens. Packing more text into an image raises the compression ratio without creating a correspondingly long text-token sequence.

Higher compression eventually makes characters too small for reliable recognition. LensVLM handles that limit with a learned tool-calling loop rather than forcing the vision encoder to decode every compressed line.

  1. The model receives compressed page images and a question.
  2. It scans the thumbnails for pages likely to contain the answer.
  3. It calls an expansion tool for selected pages.