On-device multimodal search
Google has announced a new open-weight model for developers: EmbeddingGemma 2. The 740M-parameter multimodal model maps text, images, video and audio into a single vector space, aiming to power search experiences that run directly on devices.
Why it matters
The model targets privacy-first scenarios where data never leaves the device. According to Google, this approach enables ultra-low-latency local solutions such as search-as-you-type media retrieval, keyframe video moment finding and zero-shot intent routing.
For developers
Teams can integrate the model cross-platform using MediaPipe Tasks. Those seeking finer performance tuning can optimize across CPU, GPU and NPU accelerators with LiteRT.
An open-weight approach
Releasing the weights openly gives flexibility to those building applications that run locally or on private infrastructure.
