Google DeepMind has introduced EmbeddingGemma 2, a new lightweight multimodal model designed for on-device operations. Unlike its predecessor, which focused exclusively on text, the updated version is capable of processing multiple types of input.
The original EmbeddingGemma, released in September 2025, utilized 308 million parameters and was based on the Gemma 3 architecture. It supported over 100 languages and featured a 2K token context window.
The first model was optimized for efficiency, requiring less than 200MB of RAM to run. It achieved inference speeds of 15 to 22 milliseconds on EdgeTPU hardware and allowed developers to adjust output dimensions between 128 and 768.
Embedding models convert content into numerical representations to support search, recommendations, and retrieval-augmented generation. By performing these tasks on-device, the technology ensures that sensitive data remains local.
