> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# NeoMME: an open-source model that searches scanned documents without OCR
- URL: https://globalfeed.ai/en/h-company-releases-neomme-single-tower-multimodal-encoders-at-260m-and-800m-parameters/
- Published: 2026-09-08T10:19:07.000Z
- Updated: 2026-09-08T10:19:07.000Z
- Description: H Company released two encoders, Apache 2.0, that read a PDF page as an image and search inside it. The 260M version comes within a hair of a model 14 times its size.
- Author: GlobalFeed Editor
- Tags: NeoMME, H Company, document retrieval, open source, x-hcompany_ai, dil-en, video, x-gitti

If you hold thousands of scanned invoices, contracts or technical specifications, the usual way to search them is to run OCR first. OCR loses the table, the column layout and the chart, and search cannot find what OCR threw away.

**NeoMME** skips that step. It reads the page as an image and returns the right page from a plain sentence such as "the fourth quarter revenue table from last year". H Company published two sizes, 260 million and 800 million parameters, under Apache 2.0, commercial use included.

## Small model, large result

On ViDoRe v3, the standard benchmark for document retrieval, the 260M model scores 0.523 nDCG@10\. That is ahead of every model below 800M, and only 0.002 behind ColQwen2.5 at 3.75 billion parameters, roughly fourteen times its size. The 800M version reaches 0.556.

For an operations team the two numbers that matter are different. On a single NVIDIA L40S the 260M model indexes 51.3 pages per second, about twice ColModernVBERT. And the stored vector per page drops from about 1.5 MB to 6 kB, a factor of 255, while keeping 95 percent of retrieval quality. A one million page archive needs roughly 6 GB of index instead of 1.5 TB.

## How it works

The established approach in this field bolts a pretrained vision model onto a pretrained language model; ColPali and its descendants are built that way. NeoMME uses neither. A single Transformer runs text and raw image patches cut from the page through the same layers, and the model is trained from scratch. The pretrained parts it does not carry are parameters it does not spend.

Both models support a context of 16,384 units, enough for two 4K pages. Multilingual use was part of the design: the tokenizer produces 44.4 percent fewer units than ModernBERT across 14 languages, so the same page costs less to process. Training ran on 16 and 32 H100 cards over roughly 524 billion units.

## Where it is weak

The model is strong on visual documents and not on plain text. For text-only retrieval it scores 0.4881 and 0.5126 on BEIR-15, while LateOn at 149 million parameters leads with 0.5722\. The authors attribute the gap to the scale of supervision during training. NeoMME is not a replacement for text search; it is a separate tool for scanned documents where layout carries meaning.

Models and measurements are in [H Company's announcement](https://huggingface.co/blog/Hcompany/neomme?ref=globalfeed.ai), and the method is described in the [arXiv paper](https://arxiv.org/abs/2609.01657?ref=globalfeed.ai).