01
What is Apple LensVLM-9B?
Apple publishes LensVLM-9B through its verified Hugging Face organization. It is a 9-billion-parameter vision-language model derived from Qwen3.5-9B and designed for question answering over long, text-heavy documents.
Its core technique is selective context expansion: inspect compressed page representations first, then expand only pages likely to contain the answer. Apple's reference workflow documents 5x, 10x and 15x compression settings.
02
Download, code and licensing
The official model identifier is apple/LensVLM-9B on Hugging Face, with apple-aiml-research/ml-lensvlm as the reference repository. The model weights use the Apple Machine Learning Research Model License.
The reference source code is separately licensed under Apple's Sample Code License. Publicly downloadable weights therefore should not be described as Apache-2.0 or automatically unrestricted for every commercial use; teams should review the actual terms.
03
Local serving and API options
Apple's reference repository is the safest starting point when you need the full selective-expansion document workflow. Hugging Face also documents model loading or serving with Transformers, vLLM, SGLang and Docker Model Runner.
No Apple-hosted LensVLM API was verified, and the official Hugging Face page currently shows no Inference Provider. vLLM or SGLang can expose a self-hosted OpenAI-compatible endpoint, but that endpoint is operated by you rather than Apple.
04
What the reported benchmarks show
The LensVLM paper reports 68.9% question-answering accuracy and 76.8% page-selection accuracy for the 9B configuration in one scale experiment, compared with 31.3% QA accuracy for the listed 9B base condition.
Those are author-reported research results and depend on the paper's datasets and setup. They should not be generalized into a claim that LensVLM is universally more accurate or cheaper than every OCR, parser or document-RAG system.
05
Hardware, GGUF and practical limits
Apple does not publish one simple official minimum-RAM or minimum-VRAM table in the reviewed model card. Runtime memory depends on precision, vision components, context, KV cache and the serving framework.
Community GGUF conversions exist, but they are third-party artifacts rather than Apple releases. Validate quantized output against the official BF16/reference workflow and do not turn raw file size into an unsupported hardware guarantee.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.