H Company releases NeoMME multilingual multimodal encoders

H Company released NeoMME on September 3, a multilingual encoder family that processes text and image patches within one bidirectional transformer. The announcement describes 260-million and 800-million-parameter variants, with checkpoints under Apache 2.0.

One encoder for multiple retrieval representations

NeoMME is trained from scratch using masked discrete diffusion, rather than attaching a separate pretrained vision tower. Retrieval variants produce dense and late-interaction embeddings in a single pass. The release targets workloads that include visually structured document pages without requiring an OCR stage.

This is an encoder announcement. Its outputs support tasks such as finding relevant material; they should not be confused with a model that generates an answer to a user’s question.

Design evaluation around the document collection

Our analysis: test whether the chosen representation preserves the distinctions that matter in the actual corpus. A model can retrieve thematically similar pages while missing a specific table row, language variant or visual label needed to answer a question.

Dense and late-interaction approaches also have different storage and serving implications. Measure indexing time, index size and query latency separately. If a pipeline first selects candidates using dense vectors, later reranking cannot recover a relevant page excluded from that shortlist.

For multilingual collections, inspect results by language and document type instead of relying on one combined score. Include scanned pages, mixed-language content and layouts resembling production inputs.

An OCR-free path can simplify one stage of ingestion, but it still requires decisions about page resolution, cropping and document boundaries. Preserve those preprocessing settings with the checkpoint version so that an index can be reproduced and refreshed consistently when either the model or source material changes.