H Company's NeoMME rethinks visual document retrieval

H Company has introduced NeoMME, a foundation encoder built specifically for visual document retrieval — the task of finding relevant pages, screenshots, and richly formatted documents rather than plain text passages. According to the company's write-up on Hugging Face, the model is designed to be multimodal-native and multilingual from the ground up, so images and text are handled by one shared pathway.

The headline architectural choice is a single-tower design. Many retrieval systems use separate encoders for queries and documents, which adds complexity and cost. Folding both into one tower is meant to simplify training and deployment while keeping the model versatile across languages and document types.

Why it matters

Retrieval-augmented systems increasingly need to search over what documents actually look like — tables, charts, forms, and scanned pages — not just their extracted text. NeoMME targets that gap with a few practical priorities:

• Compact late-interaction embeddings, which retain fine-grained matching between query and document tokens without ballooning storage.
• Multilingual coverage, so the same encoder can serve document search across languages.
• Efficient fine-tuning and inference, positioned as a key advantage for teams operating at scale.

Late-interaction methods, popularized by ColBERT-style retrieval, tend to be more accurate than single-vector approaches but heavier to store and serve. By emphasizing compact embeddings, NeoMME is pitching a balance between retrieval quality and operational cost.

The release is described as an initial foundation encoder under a non-standard license, so teams will want to check terms before production use. Full benchmark details and training specifics are laid out in the accompanying paper.