Apple Open-Sources LensVLM-9B: 100-Page Documents Compressed into Images First, KV Cache Reduced by 84%

Beating AI News: Apple has open-sourced LensVLM-9B, a vision-language model specifically designed for processing long documents, trained on Qwen3.5-9B-Base. When processing long documents, it first compresses the entire document into low-resolution pages for a quick scan, locates the relevant pages, and then reads the original text or high-resolution images from them.

This eliminates the need to stuff the full text into the model's context. In paper tests on a 100-page document, directly reading the full text required 51,273 tokens, while LensVLM used only 8,090. The corresponding KV cache dropped from approximately 1.6GB to 253MB, a reduction of 84.2%.

Accuracy did not drop significantly after compression. Across 7 document question-answering tests, directly reading the full text achieved an average accuracy of 72.4%, while LensVLM still achieved 68.9% at approximately 4.3x compression. Compared to directly compressing text into images for the model to recognize, at the same approximately 5x compression, accuracy improved from 31.3% to 68.9%.

However, it is currently slower. After LensVLM locates the relevant pages, it still needs to call a tool again to read the original text, requiring two consecutive rounds of inference. In paper tests, one answer takes approximately 17 seconds; without this compression scheme, directly feeding the entire document as text into Qwen3.5-9B takes approximately 8 seconds to generate an answer.