Sakhanda Wire
NVDA $153.59 -1.64% MSFT $507.00 -2.12% GOOGL $279.98 -1.32% META $619.94 -1.03% AMZN $256.62 -0.44%
← Back to the news

H Company Releases Single-Tower NeoMME Multimodal Encoders

H Company has released NeoMME, a family of 260M and 800M bidirectional multimodal encoders designed for visual document retrieval. The models use one Transformer for multilingual text tokens and raw 32×32 RGB image patches, without a separately pretrained vision tower or a causal decoder. Both support a 16,384-token context and were trained from random initialization.

NeoMME-Retriever adds dense and late-interaction retrieval heads to the shared backbone. On ViDoRe v3, the 260M model records 0.523 nDCG@10, while the 800M version reaches 0.556. The smaller model is within 0.002 of ColQwen2.5-v0.2, which has 3.75B parameters. On one NVIDIA L40S, NeoMME-260M indexes 51.3 pages per second; a query takes 78.3 ms on a CPU-only host.

The models are available under Apache 2.0 and have support in Hugging Face Transformers. Hierarchical pooling and quantization reduce storage from about 1.5 MB to 6.0 kB per page in one configuration. The authors report weaker text-only retrieval, with BEIR-15 scores below those of LateOn, and identify frozen natural-image transfer as another limitation.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news