Apple open-sources LensVLM-9B: reads long docs as compressed images
The 9B open model holds full-text accuracy at 4.3x compression and beats baselines to 10.1x. GGUF quants are up — but check Apple's license before shipping.

Copy markdown
A VLM that reads pages as pictures
LensVLM-9B renders text to images, then uses learned tools to expand only the pages a query actually needs — holding uncompressed-text accuracy at 4.3x effective compression and beating baselines out to 10.1x across seven text-QA benchmarks. For long-document RAG and OCR pipelines, that's a fraction of the vision tokens per document at the same answer quality.
Weights and GGUF quants are already up
The 9B weights — built on Qwen3.5-9B-Base — run today via Transformers, vLLM, SGLang or Docker, and the community has already shipped GGUF plus MXFP4/MXFP8 quants. That's enough to wire it into a local document-QA agent or a cheap PDF-parsing service this weekend on modest hardware.
Check the license before prod
It ships under the Apple Machine Learning Research Model License, not Apache or MIT — fine for prototyping and evaluation, but read the terms before you put it in a shipping product.