According to Google DeepMind, EmbeddingGemma is a new open embedding model built for on-device AI and described as best-in-class, aimed at deployment without cloud dependency for real-world applications (source: Google DeepMind on X, Sep 4, 2025). Google DeepMind states the model has 308M parameters and targets state-of-the-art performance while remaining small and efficient for broad hardware coverage (source: Google DeepMind on X, Sep 4, 2025). Google DeepMind adds that EmbeddingGemma can run anywhere, including without an internet connection, highlighting offline inference capability for edge devices and mobile AI workloads (source: Google DeepMind on X, Sep 4, 2025). The post does not provide public benchmarks, licensing details, or release artifacts beyond these claims (source: Google DeepMind on X, Sep 4, 2025). For trading context, the emphasis on efficient on-device and offline inference may guide attention toward edge AI workloads and mobile AI use cases referenced in this announcement (source: Google DeepMind on X, Sep 4, 2025).