Google launches EmbeddingGemma 2 for on-device multimodal search

Google has introduced EmbeddingGemma 2, an open, lightweight multimodal embedding model with 740 million parameters, according to The Verge. The update extends beyond text to capabilities across coding, images, video, and audio. Google says the model is designed for on-device tasks such as finding a video clip from a voice memo or searching audio recordings with a text prompt. Those examples position the release around retrieving information across formats.

The market significance sits in the data-preparation and retrieval layer: embedding models help make different kinds of information searchable. Google's examples suggest a broader retrieval interface in which the query and the target material need not share the same format. For builders, that makes local multimodal search a concrete product direction to evaluate. The release expands Google's offering in this layer, but the reported parameter count alone does not establish search quality, responsiveness, or practical device requirements. The brief announcement provides no comparative benchmarks to support a competitive ranking.

A concrete next step for builders is to test the two retrieval workflows Google describes against their own material. Evaluation should measure whether relevant clips and recordings are found, alongside latency and memory use on intended devices. That would help determine whether the model's broader modality coverage translates into a useful search experience. Investors should likewise distinguish the announced capabilities from evidence of adoption or commercial traction, neither of which is supplied in this report.