Google (谷歌) launched EmbeddingGemma 2, a 740-million-parameter multimodal embedding model designed for on-device privacy-focused search.

Google (谷歌) launched EmbeddingGemma 2 as a multimodal model, according to financialjuice. A separate Financial_Express roundup identifies it more specifically as an on-device multimodal embedding model with 740 million total parameters.

The reported positioning emphasizes privacy-focused search on the device. The roundup also says the model can be reduced in size, but does not describe the reduction method or provide a resulting parameter count.

The available reporting does not substantiate comparative performance claims. Although the roundup relays a claim that it is the most capable on-device multimodal embedding model, it supplies no benchmark results, supported-modality list, memory requirements, latency measurements, or licensing terms.