H has released NeoMME, a family of multilingual, multimodal encoders built for tasks such as searching visually complex documents. The two models contain 260 million and 800 million parameters and are available through Hugging Face Transformers under the Apache 2.0 license.
Unlike many visual-language systems, NeoMME does not attach a separately pretrained image encoder to a text-generating language model. One bidirectional Transformer processes both text tokens and raw image patches. The team trained the architecture from scratch with a masked discrete-diffusion objective, then adapted it for page-image retrieval using the ColPali approach.
NeoMME-Retriever produces both dense and late-interaction embeddings in one pass. At a 2048-by-2048 image resolution on an Nvidia L40S GPU, the 260-million-parameter version encoded about 51 pages per second, roughly twice the throughput reported for ColModernVBERT in the same setup. Hierarchical pooling and asymmetric quantization reduced late-interaction index storage from about 1.5 megabytes to 6 kilobytes per page while retaining more than 95% of the baseline nDCG@10 retrieval score.
Those results come from the developers’ ViDoRe v3 evaluation and specific hardware, so production speed and relevance will vary with documents and deployment choices. The release is most notable for offering downloadable checkpoints that teams can test and run themselves rather than requiring a hosted generative model for every retrieval query.