Google releases EmbeddingGemma 2 for multimodal search on-device

Google DeepMind released EmbeddingGemma 2 on October 6, 2026. The open-weight model maps text, code, images, video frames, and audio into a shared embedding space for local search, retrieval-augmented generation, classification, and related applications. Google lists a total model size of 740 million parameters and an Apache 2.0 license. (https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/; https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2)

Multimodal retrieval on local hardware

EmbeddingGemma 2 produces 768-dimensional vectors and is designed for on-device use. Its modular encoders can run text-only workloads with a smaller footprint, while optional vision and audio encoders add the other modalities. Google’s developer announcement describes offline search across local media, including text queries over images, video, and audio, without first converting every input to a text caption. The model is intended for privacy-sensitive applications where local processing and low latency matter. (https://developers.googleblog.com/google-ai-edge-with-embeddinggemma-2/)

The model also supports Matryoshka Representation Learning, which lets developers truncate embeddings to fewer dimensions to reduce vector-store size. Google’s model card describes outputs from 128 through 768 dimensions and an 8,192-token context length for text and source code. Google reports an improvement on its Code MTEB benchmark; treat those benchmark figures as the publisher’s evaluation rather than an independent performance guarantee. (https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2; https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/)

Availability and what is still coming

Google says the model weights are available through its model distribution channels and describes the model as open and commercially usable under Apache 2.0. Its October 6 developer post said an Android ML Kit service was planned; the cited announcement does not establish whether that service later became available. Check the current model card and distribution pages for supported runtimes and downloads. (https://ai.google.dev/gemma/docs/releases; https://developers.googleblog.com/google-ai-edge-with-embeddinggemma-2/)

Sources

• Google — “EmbeddingGemma 2: an open, lightweight multimodal embedding model” (October 6, 2026): https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
• Google Developers Blog — “Bring multimodal semantic search to the edge with EmbeddingGemma 2”: https://developers.googleblog.com/google-ai-edge-with-embeddinggemma-2/
• Google AI for Developers — EmbeddingGemma 2 model card: https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2
• Google AI for Developers — Gemma releases: https://ai.google.dev/gemma/docs/releases
• Google for Developers on X — official developer updates: https://x.com/googledevs