Google DeepMind has launched EmbeddingGemma 2, a 740M-parameter multimodal embedding model released under an Apache 2.0 license. According to Google, the model maps code, images, video and audio into a shared embedding space and is optimized for privacy-first use cases.
The release puts an open, permissively licensed embedding model into the hands of developers who want to build search, retrieval and classification systems without sending data to a cloud API. The Apache 2.0 license allows commercial use and modification, which matters for teams that need to run models locally.
What EmbeddingGemma 2 does
Embeddings are numerical representations that let a machine compare meaning across different types of content. By placing code, images, video and audio in the same embedding space, EmbeddingGemma 2 allows one system to search or match across formats instead of relying on separate specialized models.
Google framed the model as built for privacy-first use cases, meaning it can run in environments where data should not leave the device or local server. The 740M-parameter size keeps it small enough for on-device or edge deployment.
Open license and the wider multimodal race
The Apache 2.0 license is the headline for developers: it permits commercial use, modification and redistribution without the restrictions attached to many research-only releases. That makes EmbeddingGemma 2 a candidate for products that need custom retrieval pipelines.
The launch lands as other labs push multimodal systems in different directions. Reka AI has released a research preview of Rho-1, a 19-billion-parameter omni-model that handles text, images, video and robot control in one network, trained on 320 H100 GPUs over about three months, according to The Decoder. Rho-1 runs all modalities as tokens in a shared context window, a different design choice from a dedicated embedding model.
