mozgiH-Company introduced the multimodal encoder NeoMME
Company H-Company released NeoMME — an efficient multimodal and multilingual encoder designed for processing visual and text data. The model is developed to solve cross-modal search and classification tasks with an emphasis on high performance under limited computational resources.
The architecture of NeoMME supports work with over 100 languages, allowing its use in global applications without the need for retraining for specific linguistic regions. The encoder demonstrates high accuracy in tasks of matching images and text descriptions, surpassing previous analogs in data processing speed by 25% while maintaining comparable output quality.
Developers optimized the model for operation on standard server equipment, reducing memory requirements by 40% compared to classic architectures like CLIP. The model's weights are available for integration into third-party pipelines, simplifying the implementation of multimodal search in existing search systems and recommendation services.
Implementing NeoMME allows for reducing delays in real-time query processing, which is critical for computer vision and content analysis systems. Using this encoder gives companies the opportunity to deploy scalable multimedia processing solutions without significant investments in GPU infrastructure.
