Apple Open-Sources LensVLM-9B for Token-Efficient Long-Document QA
On September 24, Apple quietly open-sourced LensVLM-9B, a 9B-parameter vision-language model, on Hugging Face, along with its paper (arXiv:2605.07019) and code (GitHub ml-lensvlm).
Confirmed
- LensVLM-9B is fine-tuned from Qwen3.5-9B and released by Apple on Hugging Face without fanfare.
- Its method works in two steps: long documents are rendered into small compressed page images fed to the model as "visual compression" that saves context tokens and compute; a learned tool then mimics human book-flipping—scanning page thumbnails first and retrieving full text only for pages likely to contain the answer.
Why it matters
- Long-document processing is a key pain point for VLMs and LLMs, and LensVLM offers a new path to cutting context costs via image compression plus selective retrieval.
- Apple's full open-sourcing of weights, paper, and code enables community reproduction and improvement.
2026-09-24 ~ 2026-09-25 · 5 related posts
Primary sources
- Apple Open-Sources LensVLM-9B, a Qwen3.5-9B Finetune That Shrinks Long Docs into Page Images — victormustar ·
- Apple open-sources LensVLM-9B, a VLM that reads compressed text images and selectively expands relevant pages — jacek2023 ·
- Apple drops new HF model: Qwen3.5-9B finetune that turns long docs into page images to save tokens — yoobinray ·
- [source] Apple open-sources LensVLM-9B, a VLM that reads compressed text images and selectively expands relevant pages — jacek2023 · 2026-09-24
- [source] Apple Open-Sources LensVLM-9B, a Qwen3.5-9B Finetune That Shrinks Long Docs into Page Images — victormustar · 2026-09-24
- Apple quietly drops a Qwen3.5-9B finetune on Hugging Face that shrinks docs to page images — victormustar · 2026-09-24
- Apple releases LensVLM 9B, a VLM that skims long documents like a book — apolinariosteps · 2026-09-24
- [source] Apple drops new HF model: Qwen3.5-9B finetune that turns long docs into page images to save tokens — yoobinray · 2026-09-25