Google Releases EmbeddingGemma 2, a Five-Modality Embedding Model

According to an InfoQ report published on October 31, 2026, Google has released EmbeddingGemma 2, an open-weight multimodal embedding model. Built on Gemma 4 and released under the Apache 2.0 license, it maps text, code, images, video and audio into a single unified 768-dimensional embedding space, and is optimized for low-latency retrieval on-device. Google says embeddings from different modalities share the same dimensionality and can be compared directly by semantic similarity, so developers no longer need to maintain separate, incompatible embedding spaces per media type.

On the deployment side, the model lets applications load only the encoders they need, with parameter footprints scaling from 2.7B for text and code up to 7.4B for full multimodal capability, and supports MRL vector truncation to reduce storage. Google claims it significantly outperforms the previous EmbeddingGemma 1 on code search and technical content retrieval, and supports tasks such as finding documents, screen recordings and audio clips from a query, searching source code with natural language, and retrieving video content from text descriptions. It is pitched for on-device search, RAG, classification and clustering workloads. As of the initial announcement, no further follow-up developments have been reported.