Google has introduced EmbeddingGemma 2, a 740-million-parameter multimodal embedding model designed to run entirely on devices such as smartphones and Raspberry Pi boards, mapping multiple data types into a unified space while preserving user privacy.
- Runs completely on-device with no data sent to the cloud
- Supports over 100 languages and five content modalities
- Open-source with Apache 2.0 license but lacks safety tuning
What happened
Google has released EmbeddingGemma 2, a new open multimodal embedding model that operates fully on devices like phones and small computer boards. This model, with 740 million parameters, can embed text, code, images, audio, and video into a single 768-dimensional space, enabling meaning-based search and retrieval without internet connectivity.
Running on hardware such as a Pixel 11 Pro or Raspberry Pi, it requires between 191MB to 567MB of memory depending on the modalities utilized. Licensed under Apache 2.0, the model is an evolution of the original EmbeddingGemma and shares its core components, allowing for efficient combined usage on limited-memory devices.
Why it matters
EmbeddingGemma 2’s on-device architecture addresses significant privacy and regulatory concerns by eliminating the need to transfer personal or sensitive data to external servers. This aligns with the growing demand, especially in Europe, for local data processing and greater control over user information without compromising functionality.
By supporting over 100 languages and multiple data formats simultaneously, the model enhances accessibility and usability across diverse applications in search, recommendation, and multimodal analytics. Though safety tuning and output moderation are absent, the training data has been mitigated to reduce risks, signaling a shift towards transparent and user-controlled AI deployments.
What to watch next
Future developments will likely focus on improving model performance uniformly across languages and adding safety mechanisms to address potential misuse or harmful outputs. Monitoring how developers and regional operators, especially in privacy-conscious markets, adopt EmbeddingGemma 2 will provide insight into on-device AI’s role in the broader ecosystem.
Additionally, tracking integration in consumer devices and applications leveraging this open model will show its practical impact. Competitors and open-source communities may also respond by advancing similar or complementary on-device multimodal embedding technologies under permissive licenses.