News · 2026-09-24

Apple releases LensVLM-9B, a document model that zooms into selected pages

Apple has released LensVLM-9B weights and code for a vision-language model that sees a compressed map of a long document, then asks to enlarge only the pages needed to answer a question. The method is a practical alternative to reading every PDF page at full resolution, but the public release is research-only and Apple does not state a consumer-GPU requirement.

Key facts

• LensVLM is a 9B-parameter BF16 vision-language model with public weights and official code.
• It expands selected pages as source text, OCR text or high-resolution imagery.
• Apple reports 41.9 versus 31.1 for a compressed-image baseline on MMLongBench-Doc at 15× nominal compression.
• Primary source: Apple's LensVLM-9B model card.

The mechanism is called selective context expansion. A document becomes a sequence of small page images that preserves layout and rough content but is too blurry for exact reading. From that overview, the model predicts which page matters and calls Expand(k). The tool re-presents information already in the document as native text, OCR or a high-resolution image. The paper emphasizes that expansion does not fetch new knowledge; it makes selected existing content legible.

Think of scanning a book by flipping through thumbnails, then opening the two pages whose diagrams and headings look relevant. That differs from classic retrieval-augmented generation, which converts a corpus into chunks and fetches a few likely text passages. Retrieval can be superior when the same corpus is asked about repeatedly and an index is cheap to amortize. LensVLM keeps a global visual map, which can be better for forms, charts, layout-dependent evidence and questions whose relevant pages are hard to name before reading.

Apple reports that LensVLM remains comparable to a full-text upper bound at 4.3× effective compression and beats its retrieval, text-compression and visual-compression baselines up to 10.1× effective compression across seven text-QA benchmarks. “Effective” matters: it counts the overview plus returned expansions, so a nominal 15× compression setting is not 15× free context. In one ablation, a second expansion added 6.1 percentage points after the first had captured most of the gain. These are author-reported results rather than independent replication.

The model is downloadable, but Apple does not give a disk-size figure in the checked model card, so this article does not guess one. The official repository has Transformers, vLLM, SGLang, Docker and demo instructions. Apple does not publish a minimum inference VRAM figure. Its paper used eight B200 GPUs for certain efficiency measurements, which is a measurement configuration—not a requirement for ordinary inference.

The legal caveat is clear. Apple's Machine Learning Research Model License limits the weights to research purposes, excluding commercial use, product development and commercial services. That makes “open weights” a distribution fact, not a commercial-open-source guarantee; see what open weights means.

LensVLM matters because long-context AI often faces a choice between expensive fidelity and brittle retrieval. A model that can cheaply locate evidence before spending tokens to read it may make document agents more reliable. The unresolved questions are compute, page-selection accuracy outside the paper's benchmarks and whether independent users can reproduce the reported compression-quality trade-off.

Primary source, verified: read the paper → (arXiv 2605.07019)

Ground Truth. (2026, September 24). Apple releases LensVLM-9B, a document model that zooms into selected pages. Ground Truth. https://groundtruth.day/news/apple-lensvlm-9b-selective-document-context-expansion.html

@misc{groundtruth:apple-lensvlm-9b-selective-document-context-expansion,
  title  = {Apple releases LensVLM-9B, a document model that zooms into selected pages},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/apple-lensvlm-9b-selective-document-context-expansion.html}
}

Topics: apple · vision-language-models · documents · open-weights · research