Apple发布LensVLM-9B视觉语言模型

Apple推出了LensVLM-9B，这是一个基于Qwen3.5-9B-Base的90亿参数视觉语言模型12。该模型通过扫描压缩图像并选择性地扩展相关部分，使视觉语言模型能够在处理压缩文本图像时保持高精度1。在4.3倍有效压缩率下，LensVLM的精度与完整文本基准相当1，相比基线最高提升10.1倍有效压缩1。该模型支持5x、10x、15x三种压缩选项2，在七个文本QA基准任务上的性能超越检索和压缩基线，应用范围涵盖文本QA、多模态文档理解和代码理解等领域1。

LensVLM-9B已在Hugging Face开源2，并提供了多种推理框架的使用方法，包括Transformers、vLLM和SGLang2。代码仓库地址为https://github.com/apple-aiml-research/ml-lensvlm[2](#source-2)。相关研究论文《LensVLM: Selective Context Expansion for Compressed Visual Representation of Text》已发布在arXiv（编号2605.07019）12。

Apple has released LensVLM-9B, a 9-billion-parameter vision language model designed to efficiently process compressed text images while maintaining high accuracy 12. Built on the Qwen 3.5-9B-Base foundation and modified by Apple, the model employs a novel approach that compresses images and selectively expands only the relevant portions for processing 12.

The framework achieves performance comparable to full-resolution text benchmarks at a 4.3x effective compression rate 1. LensVLM-9B supports three compression options—5x, 10x, and 15x—and has been made available as open source on Hugging Face 2. The model excels across seven benchmark tasks, including text question-answering, multimodal document understanding, and code comprehension, outperforming retrieval and compression baselines 1.

Apple provides multiple inference framework options for deployment, including Transformers, vLLM, and SGLang 2. The corresponding research paper, titled "LensVLM: Selective Context Expansion for Compressed Visual Representation of Text," is available as arXiv preprint 2605.07019 1. The model's code repository is accessible at https://github.com/apple-aiml-research/ml-lensvlm 2.