Apple's LensVLM-9B reads text as images at 4.3× compression
Apple quietly put a 9B vision-language model on Hugging Face this week that reads documents as pictures instead of tokens — and holds its accuracy down to a quarter of the original text. It is the most interesting long-context idea to come out of a big lab this month, and it landed with almost no announcement.
Apple released LensVLM-9B on September 21 — weights under its own research licence, code under its sample-code licence, and a paper that has been sitting on arXiv since May. The pitch: stop tokenizing long documents. Render the text as page images, let the vision encoder crush each page into a fixed number of visual tokens, and use rendering resolution as a compression knob. The catch has always been that characters shrink below what the encoder can resolve, so accuracy collapses as you squeeze. LensVLM's fix is to give the model learned tools — it scans the compressed pages first, then expands only the pages it judges relevant back to full resolution, and answers from those. Built on Qwen3.5-9B, it holds accuracy comparable to feeding the model the full text at 4.3× compression, and still beats retrieval-based and text- or image-compression baselines at up to 10.1× across seven text-QA benchmarks, with the gap widening the harder you compress. It also carries over to document and code understanding.
Why it matters: long context is the industry's most expensive problem, and nearly everyone else is attacking it inside the KV cache — pruning, quantizing and sparsifying attention so a model can hold more tokens for less memory. Apple's framing is different: if text is just an image, the compression ratio becomes a rendering choice rather than an architecture change. The paper's practical finding is worth as much as the benchmark numbers — text expansion wins for rendered text, while high-resolution image expansion wins for native documents, where layout carries the answer. That is a usable rule for anyone building a document pipeline.
The honest caveats are about attention, not quality. The paper is four and a half months old; the news is the release, not the research. Engagement is thin — a couple of dozen points on Hacker News, 70 likes on the model — and the licence is Apple's own, not an OSI-approved open-source one, so "open weights" deserves the qualifier. The community has already started converting it, with GGUF builds appearing within a day, which is the signal that matters more than the star count. Apple's on-device AI story has mostly been about the Neural Engine; this is the company publishing a research idea that can be used anywhere.
Bessemer Venture Partners closed $5.75 billion across new funds on Wednesday, and $4 billion of it is earmarked for growth-stage AI. The San Francisco firm says it has invested in more than 260 AI-native companies since 2022, putting over $3 billion into the stack, with Anthropic, Perplexity, Cognition and Waymo already in the portfolio. Partner Byron Deeter said AI-native companies are "scaling faster than any category of technology we've backed before," and that companies staying private longer is "a permanent structural shift." The vehicle math is more interesting than the headline: a $3.4 billion flagship (down from $3.85 billion), a $1.85 billion growth fund (more than double its predecessor), and $500 million raised in separate vehicles with the firm's limited partners. The growth team will spread its $4 billion across all three and is targeting roughly two dozen companies — a bet that the biggest AI returns now sit late, not early.
OpenAI is letting nine mathematicians referee the maths its internal model claims to have solved — and the group's own announcement says it has no power over the pace. The Advisory Group on Mathematics and Artificial Intelligence, hosted at Princeton's Institute for Advanced Study, came together after OpenAI approached some of its members; they chose an independent body over an OpenAI board, are unpaid, and publish their recommendations publicly. Its current task is advising OpenAI on how to release what its internal model produced: more than 100 long-standing open problems, plus a claimed Navier–Stokes Millennium Prize solution, from a model OpenAI says it began training on August 28. OpenAI's announcement is explicit that the group "will not be responsible for advising us on how to pace our internal progress" — nine of the field's biggest names, including Edward Witten, Timothy Gowers and Martin Hairer, holding a release lever and not a brake.
What to watch: whether LensVLM's selective-expansion trick shows up in a shipped product, and whether the maths group publishes its first recommendations before OpenAI publishes the proofs.
If a 9B model can read a compressed page image more cheaply than the text, why is anyone still tokenizing documents? Tell us in the comments.
Sources: ml-lensvlm (GitHub)
