H Company introduced NeoMME on Sept. 3, 2026, as two AI models that work in many languages, with 260 million and 800 million parameters, the numbers a model tunes as it learns. They are built for search, not for writing text: they turn text and images of document pages into a form a search system can compare. NeoMME does this with one shared model rather than pairing a separate image model with a text model, which gives teams that build document search one compact system for both.
The single model takes in both text and small squares cut from the raw page image, and it is not built on any existing image or text model. H Company trained both sizes from scratch by hiding parts of the text and having the model fill them in, using the matching image as a clue when there was one.
A search version, NeoMME-Retriever, was trained further to find document pages. H Company says the 260-million-parameter version nearly matches ColQwen2.5, a rival model about 14 times larger, on a standard test of finding the right page.
H Company also says the smaller search model handles about 51 pages per second on one Nvidia L40S chip, roughly twice as fast as ColModernVBERT in the same comparison. Two compression techniques, the company says, cut the storage each page needs in a search index from about 1.5 MB to 6 kB, 255 times smaller, while keeping more than 95% of the original search quality.
Both models are released under the Apache 2.0 license, which lets anyone use, change and share them free of charge. NeoMME was added to the Hugging Face Transformers software library on Aug. 31, 2026.
No one outside H Company has yet published tests of the search, speed or storage results. It is also not clear how well NeoMME works language by language, or whether it is ready for heavy use in real products.