Apple Releases LensVLM-9B, a 9B Vision-Language Model That Selectively Decompresses Text Images
Apple has released LensVLM-9B, a 9-billion-parameter vision-language model fine-tuned from Qwen3.5-9B-Base that processes documents as compressed images, selectively expanding only relevant pages to full resolution. The model supports 5x, 10x, and 15x compression ratios and is available under Apple's Machine Learning Research Model License.
Apple Releases LensVLM-9B
Apple has published LensVLM-9B, a 9-billion-parameter vision-language model designed to process long documents by scanning compressed image representations of text and selectively expanding only the sections deemed relevant to a query.
The model is a fine-tune of Alibaba's Qwen3.5-9B-Base, adapted by Apple's machine learning research team. Rather than ingesting raw text tokens or full-resolution page images, LensVLM first views documents as compressed visual representations, then uses learned tools to selectively "uncompress" specific pages back to their original form when the model determines they're needed to answer a question.
According to Apple, the model supports three compression ratios — 5x, 10x, and 15x — allowing users to trade off between processing efficiency and detail retention. The approach is detailed in an accompanying paper, "LensVLM: Selective Context Expansion for Compressed Visual Representation of Text" (arXiv:2605.07019), authored by a team including Roy Xie, Dan Friedman, Donghan Yu, and others at Apple.
What's confirmed
- Parameters: 9 billion
- Base model: Qwen/Qwen3.5-9B-Base
- Format: Safetensors, BF16 tensor type, includes a chat template
- Task category: Image-Text-to-Text
- Compression modes: 5x, 10x, 15x (user-selectable)
- Downloads: 1,432 in the past month on Hugging Face
- Quantized variants: 9 additional quantized models available
Apple has not disclosed a context window size, benchmark scores, or pricing for LensVLM-9B. The model is not currently deployed by any inference provider on Hugging Face, meaning developers must run it themselves via the accompanying GitHub repository (apple-aiml-research/ml-lensvlm).
Licensing
The model weights, including Apple's modifications to the underlying Qwen architecture, are released under the Apple Machine Learning Research Model License — a research-use license rather than a permissive open-source license. The accompanying inference code is distributed separately under the Apple Sample Code License. This means commercial deployment terms differ from typical open-weight releases and developers should review both licenses before production use.
Usage
Apple provides a demo script for running inference against custom documents:
python demo.py --model apple/LensVLM-9B --text_file document.txt --question "What is the main finding?" --compression 10x
Full setup instructions, data preparation steps, and evaluation scripts are available in the GitHub repository's README.
What this means
LensVLM targets a specific bottleneck in document AI: processing long text-heavy documents without paying the full token or compute cost of high-resolution vision encoding for every page. By compressing documents into images and selectively expanding only relevant sections, Apple claims to reduce the computational overhead typically required for long-document understanding — though no efficiency or accuracy benchmarks have been published to substantiate the tradeoffs at each compression level.
The research-only license signals this is an academic release rather than a product-ready model for Apple's consumer stack. Builders evaluating it should treat the lack of disclosed context window and benchmark data as a gap to independently verify before adopting the approach for retrieval-augmented or long-document pipelines.
Related Articles
Fish Audio releases Drama 3, a 35B-parameter MoE text-to-speech model with 16-speaker support
Fish Audio has released Drama 3, a text-to-speech model with 35B total parameters and 3B active per token, built for multi-character, long-form audio. It supports up to 16 speakers per text, voice cloning, and streaming. Fish Audio reports first audio in under 100 ms on a single H200.
Apple brings Siri AI beta to Apple TV 4K 2nd and 3rd generation, alongside new $299 A19 model
Apple says Siri AI will be available as a beta on Apple TV 4K (2nd generation or later) through a forthcoming tvOS update. The company also introduced a fourth-generation Apple TV 4K with an A19 chip and a $299 starting price.
OpenBMB releases MiniCPM-V 4.7 1B, an on-device vision model with 4x/16x visual token compression
OpenBMB has released MiniCPM-V 4.7 1B, a 1B-parameter multimodal model for single-image, multi-image and video understanding on phones, tablets and local hardware. OpenBMB claims it substantially outperforms Qwen3.5 0.8B in overall score and inference efficiency in Instruct mode. Benchmark figures, context window and training cutoff are not given in the available model card text.
Microsoft releases FrogNano-4B, an Apache 2.0 coding agent trained with RL on 1,500 synthetic tasks
Microsoft has released FrogNano-4B-2609, a repository-level coding agent derived from Qwen3.5-4B and published under Apache 2.0 with open weights. Microsoft says it was post-trained only with reinforcement learning on about 1,500 synthetic software-engineering tasks, with no stronger-model trajectories. It is evaluated at roughly 131K tokens of context.
Comments
Loading...