Google releases EmbeddingGemma 2 open multimodal model for on-device applications

Google DeepMind has released EmbeddingGemma 2, an open multimodal embedding model designed to run locally on mobile devices and smartphones. SiliconANGLE, MarkTechPost, and The Next Web all reported that the new model expands the original EmbeddingGemma architecture beyond text processing. MarkTechPost noted that the model is released under the Apache 2.0 open license. [1][2][3][4]

The release builds on the Gemma 4 architecture and contains 740 million parameters, according to MarkTechPost. The outlet reported that the model maps five distinct input types into a single 768-dimensional embedding space. According to BigGo Finance, the model is engineered to run on device using 567 megabytes of memory. [2][6]

By bringing images, audio, and video into a shared embedding space with text, the architecture enables cross-modal search on local hardware, SiliconANGLE reported. For instance, an application can match a voice recording directly to a specific timestamp in a video clip without uploading data off the device. The Decoder added that Google claims EmbeddingGemma 2 outperforms rival embedding models that are twice its size. [1][5]

The original text-only EmbeddingGemma model was introduced by Google in September 2025, SiliconANGLE reported. EmbeddingGemma models use techniques such as geometric embedding distillation from larger systems and spread-out regularization to remain efficient under quantization, allowing deployment on resource-constrained platforms including mobile phones, laptops, and desktop computers [10]. [1][10]