LensVLM-9B
9B document VLM that scans compressed pages and expands only what matters.
About
LensVLM-9B is Apple's vision-language model for long documents, published in September 2026. It is a fine-tune of Alibaba's Qwen3.5-9B-Base with a new inference idea: instead of processing every page at full resolution, it first scans the document as compressed images, then selectively expands only the relevant pages back to full detail using learned tools. Users pick 5x, 10x, or 15x compression to trade efficiency against detail.
Apple published the paper (arXiv:2605.07019) and code but disclosed no context window size, benchmark scores, or pricing, so the accuracy cost of compression is currently unmeasured. The weights are on Hugging Face under Apple's Machine Learning Research license, which is limited to non-commercial scientific research. No inference provider hosts it, so developers self-host via the official GitHub repo.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No published benchmarks; Apple disclosed no benchmark scores, context window, or pricing as of September 2026 (TPS Report); compression tradeoffs at 5x/10x/15x are author-described but unmeasured
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics, metrics-v1). Evidence: official apple-aiml-research/ml-lensvlm repo (created 2026-09-22), HF apple/LensVLM-9B with license text, paper arXiv:2605.07019. Strengths: novel selective-expansion approach from Apple research, open weights and code. Docked for: (1) zero published benchmarks or context-window disclosure; (2) Apple ML Research license bars commercial use. License: Apple Machine Learning Research Model License (non-commercial).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 4.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Apple, published Sept 2026; Qwen3.5-9B-Base fine-tune; paper arXiv:2605.07019
- · Selective 5x/10x/15x visual compression; no published benchmarks or context window