Apple ships LensVLM-9B: an open VLM that reads compressed docs
A rare open-weight drop from Apple: 9B on a Qwen3.5 base, 262K context, FP16 on one 4090, GGUF quants already up — and a license to read first.

Copy markdown
Reads documents as compressed images — and expands only what it needs
LensVLM-9B takes text as compressed visual tokens (5x, 10x or 15x) and uses learned tools to expand only the pages a query actually touches, so long PDFs stop blowing up your token count. It's a 9B model on a Qwen3.5 base, BF16, with a 262K-token context.
Runs on one GPU, and the quants already shipped
FP16 fits in ~19GB at short context — a single RTX 4090 — while the full 262K window wants ~53GB (one A100 or three 4090s). It loads in Transformers, vLLM and SGLang, and community GGUF and EXL3 quants are already on the Hub, so you can pull it to a local box tonight.
Why an open release from Apple is the real story
Apple almost never publishes weights you can actually run; LensVLM lands under the Apple ML Research license — usable, but not Apache or MIT, so check the terms before you build a product on it. For local doc-QA and RAG over scans, it's a credible way to drop a per-page cloud OCR bill.
Prefer an API? DeepSeek-V4.1-Flash is the cheap doc workhorse
If self-hosting isn't your thing, DeepSeek's V4.1-Flash (MIT, native vision, 1M context) hits DocVQA 95.6 at about $0.15 per million input tokens on a cache miss — a low-cost hosted option for the same document-parsing jobs.