> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Apple releases LensVLM-9B, a research model that reads compressed images of text
- URL: https://globalfeed.ai/en/apple-releases-lensvlm-9b-a-research-model-that-reads-compressed-images-of-text/
- Published: 2026-09-27T22:52:09.000Z
- Updated: 2026-09-27T22:52:09.000Z
- Description: Apple has published LensVLM-9B on Hugging Face, a model that scans long texts as shrunken page images and makes only the page it needs readable again. The model files can be used for non-commercial research only.
- Author: GlobalFeed Editor
- Tags: Apple, LensVLM-9B, vision-language models, research, x-Apple, dil-en, elle

**Apple** published a vision-language model called **LensVLM-9B** on Hugging Face in September 2026\. According to the model card, LensVLM scans compressed images of text and expands only the pages that matter for the question back to their readable form. The paper behind the model, by researchers from Apple and Duke University, was submitted to arXiv on May 7, 2026.

## What LensVLM is

Turning a long text into images and having a vision-language model read them lets the same content be processed with less input. According to the paper, as compression increases the characters shrink below the resolution the model can see, and accuracy drops quickly. LensVLM approaches this with selective expansion. The name comes from that idea: like a lens, it magnifies the region it picks out of the compressed content.

## How it works

The text is first rendered into a sequence of page images. The model sees all the images and the question, then calls a tool it has learned to use and asks for a chosen page to be expanded. Depending on the task, the tool returns the original text of that page, text extracted from the image, or a high-resolution image. One page is expanded per step. The tool adds no new information to the document; it presents the selected part again in readable form. In the example in the repository, the model finds the relevant page in a 15-page compressed document and answers in two steps.

LensVLM-9B is built on Alibaba's Qwen3.5-9B. According to the notice published with the model card, the architecture is unchanged; Apple retrained the model files for the selective expansion task. Training has two stages: first on synthetic examples that show tool use, then with reinforcement learning. The repository comes with three compression settings.

## Who can use it

The model files are released under the Apple Machine Learning Research Model License. According to the license text, use is limited to non-commercial scientific research and academic work; use in a commercial product or service is outside its scope. The code is distributed under a separate license, the Apple Sample Code License. The model is run with the sample scripts in the GitHub repository; no minimum hardware requirement is stated.

According to the authors' own measurements, LensVLM keeps accuracy close to full text on question answering tests while reading text compressed several times over, and it generalizes to code and document understanding tasks. The same paper gives the limits as well: on the document task at low compression, the base model stays ahead of the versions that use the tool, and on code tasks the gain is smaller because the texts are short. According to the authors, the method helps most with long texts. The model card has no separate section on usage limits or supported languages. To see how it performs on other tasks, it is worth following evaluations from independent researchers as well.

Sources: [Hugging Face: apple/LensVLM-9B](https://huggingface.co/apple/LensVLM-9B?ref=globalfeed.ai), [arXiv: LensVLM paper](https://arxiv.org/abs/2605.07019?ref=globalfeed.ai), [GitHub: ml-lensvlm](https://github.com/apple-aiml-research/ml-lensvlm?ref=globalfeed.ai).