LensVLM: Compressing long context as images, expanding only relevant pages

LensVLM-9B is a vision-language model that encodes very long textual contexts by compressing pages of text into visual representations and then selectively expanding only the pages relevant to a given query. The approach - described in arXiv:2605.07019 as "Selective Context Expansion for Compressed Visual Representation of Text" - trades dense token-level context for a compact visual summary, enabling long-context multimodal understanding while keeping inference costs lower by decompressing just the pertinent parts. The model is built on a Qwen‑3.5‑9B base (finetuned) and is targeted at tasks requiring image-text-to-text reasoning, conversational multimodal interactions, and visual text compression.

The release is provided as a Hugging Face model repository (safetensors, ~9.4B parameters, ~18.8 GB total files) under the apple‑amlr license, with example code for Transformers pipelines and direct AutoModelForMultimodalLM use (processor.apply_chat_template, model.generate) as well as vLLM serving and an OpenAI‑compatible curl example. Tags highlight long-context, vision-language, and conversational use cases; the page lists downloads, likes, and compatibility details (device_map support, inference endpoints). The package includes processor and pipeline integration, making it practical to run compressed-image context prompts and query-selective expansion workflows out of the box.