Report says Apple has open-sourced a long-document vision-language model LensVLM-9B

AT A GLANCE

LensVLM-9B Compressing long documents first and then reading the relevant pages can substantially reduce KV cache usage, but responses take longer.

Article

Apple open-sourced LensVLM-9B：100 For documents, compress pages into images first,KV saving cache 84%Dongcha Beating AI news: Apple has open-sourced a vision-language model specifically designed to process long documents LensVLM-9B, trained on Qwen3.5-9B-Base. When processing long documents, it first compresses the entire document into low-resolution pages for a quick scan. After locating the relevant pages, it then reads the original text or high-resolution images within them.

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

Accuracy also did not decline significantly after compression. 7 In 72.4%document question-answering tests, the average accuracy when directly reading the full text was 4.3. Even with approximately 68.9%× compression, LensVLM still achieved 5. Compared with directly converting the text into images for the model to recognize, at the same approximately 31.3%× compression, accuracy increased from 68.9%to 17.

However, it is currently slower. After LensVLM locates the relevant pages, it must make another tool call to read the original text, and therefore requires two consecutive rounds of inference. In the paper's tests, a single response took approximately 8 seconds; without this compression approach, directly inputting the entire document as text into Qwen3.5- required approximately https://m.theblockbeats.info/flash/368846

Key points

Report says Apple has open-sourced a long-document vision-language model LensVLM-9B。
LensVLM-9B Based on Qwen3.5-9B-Base training.
Processing 100-page documents, LensVLM uses 8090 units tokens, while directly reading the full text uses 51273 units tokens。
corresponding KV cache from approximately 1.6GB to 253MB, reducing 84.2%。
7 document question-answering tests, the average accuracy of directly reading the full text was 72.4%, while LensVLM achieved 68.9%。
paper tests, a single LensVLM response takes approximately 17 seconds, while directly inputting the full text takes approximately 8 seconds.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

LensVLM-9B It first uses low-resolution pages to quickly locate relevant content, then reads the original text or high-resolution images of the relevant pages, thereby reducing the amount of content that needs to be placed in the context. The trade-off is that two consecutive rounds of inference are required.

Why it matters to readers

Reported tests show that this method can reduce, when processing long documents,tokens and KV cache usage while maintaining accuracy close to that of directly reading the full text.

It can be understood as quickly browsing the entire document first, then closely reading the relevant pages, thereby reducing the amount processed at one time.

Risks and unknowns

The current evidence comes from media reports and does not provide an official Apple announcement, code repository, or original paper link.
The open-source attribution, test results, and latency data still require confirmation through primary materials or independent reproduction.
The evidence did not state whether the model had been incorporated into Apple products or developer tools or formed part of a commercial arrangement.

Report says Apple has open-sourced a long-document vision-language model LensVLM-9B

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center