01
Start with Apple's reference implementation
The official model is apple/LensVLM-9B. When the goal is to reproduce LensVLM's selective document expansion rather than simply load the base VLM, begin with Apple's ml-lensvlm repository and its documented demo workflow.
The reference workflow accepts a document, question and compression setting. Apple documents 5x, 10x and 15x modes, which makes the reference implementation the right baseline before optimizing for a different serving stack.
02
Transformers, vLLM and SGLang
Hugging Face documents LensVLM-9B with its image-text-to-text Transformers interfaces. It also shows vLLM serving and an OpenAI-compatible chat-completions path for applications that already use that protocol.
SGLang is another documented serving option, and Hugging Face also exposes a Docker Model Runner path. These are self-hosted integrations; none of them creates an Apple-managed LensVLM API.
03
Community GGUF conversions are not official Apple releases
Community members have published LensVLM-9B GGUF conversions at multiple quantization levels. Those files can be useful for memory-constrained experiments but are separate third-party artifacts with their own compatibility and quality considerations.
A GGUF file's raw size is not the full runtime requirement. Multimodal projection handling, context, KV cache and framework overhead can add memory, and some conversions may not reproduce Apple's complete selective-expansion workflow.
04
Hardware guidance without invented minimums
Apple does not publish one universal minimum VRAM or RAM requirement for LensVLM-9B in the reviewed launch material. BF16 weights are substantially larger than low-bit community quantizations, but actual serving memory depends on the runtime and workload.
Benchmark the exact model representation, context length, image/document workload and server framework. Avoid claims such as runs on any 8 GB GPU unless a tested configuration and its limitations are documented.
05
Which local route should you choose?
Use Apple's repository when you need the research workflow and selective page expansion. Use vLLM or SGLang for an application server, and use Transformers for direct Python integration.
Consider community GGUF only when you are prepared to validate multimodal behavior and quality yourself. If a generic loader works but the document workflow behaves differently from the paper demo, compare it against Apple's reference implementation before diagnosing the model.
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.