Show HN: We beat Gemini Embedding 2 by training only 16M params (open weights) (huggingface.co)

A new multimodal embedding model, Fusion Embedding 1, has been introduced, showcasing significant advancements in cross-modal retrieval capabilities by training only 16 million parameters while extending an existing state-of-the-art vision-language embedding system with an audio modality. This model builds upon the Qwen3-VL-Embedding-2B framework but does so without altering the base model's weights. Instead, a trained connector maps audio-tower features into the existing embedding space, resulting in a cohesive framework that allows for effective retrieval of text, images, video, and audio across modalities. The significance of Fusion Embedding 1 lies in its performance, surpassing established models like ImageBind, LanguageBind, and Gemini Embedding 2, particularly in audio-to-text retrieval. The model achieves robust emergent alignment between audio and image retrieval despite being trained exclusively on audio–text pairs, showcasing its potential in practical applications. The compact size of the trained connector and the ease of integrating it with existing architectures make Fusion Embedding 1 an attractive option for researchers in the AI/ML community looking to explore scalable, efficient models capable of handling diverse data types. The repository is freely available, promoting further experimentation and enhancement in multimodal embedding research.