A new open-source AI model just made document search 255x cheaper to store

Hcompany has released NeoMME, an open-weights, Apache 2.0 encoder for searching inside documents. It is available now in Hugging Face Transformers, in 260M and 800M sizes.

NeoMME turns pages of text, PDFs, screenshots, and scanned documents into vectors a search engine can rank, the "search inside documents" layer behind many AI assistants. It does the work with a single bidirectional transformer that reads both text tokens and raw image patches. There is no separate pretrained vision tower and no causal language model bolted on. The closest open comparator, ColModernVBERT, still uses a separate SigLIP2 vision encoder in front of its text stack, which means more parts to host and fine-tune.

NeoMME's late-interaction index, the per-page vector footprint a retrieval system has to keep on disk, drops from roughly 1.5 MB to about 6 kB per page, about 255x smaller, while retaining more than 95% of baseline nDCG@10 on ViDoRe v3. Throughput is also up: at a matched 2048×2048 input on an NVIDIA L40S GPU, the 260M model encodes about 51 pages per second, roughly twice ColModernVBERT, per the Hcompany blog.

For builders, this is a cheaper default for document-RAG infrastructure. Small teams can now afford to index larger corpora on a single box. The honest read: an incremental efficiency release inside a crowded open-weights retrieval field, not a frontier claim, and the accuracy-retention figure should be re-checked against the companion arXiv paper at matched precision and batch conditions.