NeoMME Is a Tiny AI Model That Reads Documents Twice as Fast as Its Rivals
A new open-source model scans document pages as images, skips the usual heavy architecture, and still matches much larger systems on accuracy. Here is what that means if you work with PDFs.
Key points
- NeoMME comes in two sizes, 260 million and 800 million parameters, and Hugging Face released both under a free Apache 2.0 licence on 22 May 2025.
- The smaller 260M model processes roughly 51 document pages per second on a single NVIDIA L40S GPU, about twice the speed of the comparable ColModernVBERT model.
- NeoMME-Retriever-260M scores 0.523 on the ViDoRe v3 retrieval benchmark while using around 14 times fewer parameters than ColQwen2.5, which scores only 0.002 points higher.
- A compression technique cuts the storage each indexed page needs from roughly 1.5 megabytes down to 6 kilobytes, a 255-fold reduction, while keeping more than 95 percent of retrieval accuracy.
- Both model sizes are available now through Hugging Face Transformers with no licence fee.
Searching through a pile of PDFs is one of those tasks that sounds simple but quietly wastes hours. You need the computer to read each page, understand what is on it, and then find the right one when you search for something. Most AI systems built for this job are large, slow, or expensive to run. NeoMME is a new attempt to change that.
What exactly is NeoMME?
NeoMME, pronounced "nee-oh-me", is a multimodal encoder, meaning software that turns both text and images into compact numerical summaries a computer can compare and search. Most rival systems bolt a separate image-reading component onto a text-reading component. NeoMME runs everything through a single shared engine, which trims the overall size and speeds things up.
The team trained it from scratch rather than adapting an existing large model. It comes in two sizes: 260 million parameters (think of parameters as the tunable knobs inside a model that determine how smart it is) and 800 million parameters. For reference, the model behind basic ChatGPT responses runs into the tens of billions of parameters, so both NeoMME variants are genuinely compact.
How does it read a document page?
Instead of running optical character recognition, the software process that converts a scanned image into typed text, NeoMME looks at each page as a picture. It chops the image into small 32-by-32-pixel squares and feeds those squares directly into its single processing engine alongside any search query you type.
That approach preserves things a text extractor would miss: the layout of a table, the font size used for a heading, a bar chart. A perfectly accurate text extractor still cannot tell you that a number appeared inside a red box, but NeoMME can, because it never threw the visual information away.
Why does the storage shrink so dramatically?
When a retrieval system, the part of an AI tool that finds relevant documents from a large collection, stores a detailed fingerprint of each page for later searching, those fingerprints add up fast. NeoMME uses two tricks called hierarchical token pooling (grouping nearby image patches together before storing them) and asymmetric quantization (representing numbers with less precision where full precision is not needed). Together they squeeze each page's fingerprint from about 1.5 megabytes to 6 kilobytes without meaningfully hurting accuracy.
For a business storing 100,000 document pages, that difference is between roughly 150 gigabytes of index storage and under 600 megabytes.
What does this mean for ordinary users?
If your workplace uses an AI tool to search contracts, invoices, research reports, or any other PDF-heavy archive, NeoMME is the kind of model that could sit underneath it. Faster encoding means results arrive quicker. Smaller storage means the tool costs less to run. Neither of those things requires you to do anything differently; they matter to whoever builds or buys the software you use.
If you are a developer or a small-business owner building your own document search tool, NeoMME is free to download and use commercially right now through Hugging Face Transformers. The Apache 2.0 licence means you can use it in a paid product without royalties.
The honest takeaway: NeoMME looks genuinely efficient for its size, and the benchmark numbers are promising. But every benchmark tests a curated set of documents. Before you trust it with your own archive, run it against a sample of your actual files, because real-world collections are messier than any leaderboard.
