Welcome EmbeddingGemma, Google's new efficient embedding model

Google has released EmbeddingGemma, a highly efficient, multilingual embedding model perfect for on-device applications like mobile RAG pipelines. With a compact size of just 308M parameters and a 2K context window, it supports over 100 languages while maintaining a very small memory footprint. The model's architecture uniquely uses Matryoshka Representation Learning, allowing its output vectors to be flexibly truncated for faster processing and lower memory use. Despite its small size, EmbeddingGemma delivers state-of-the-art performance on multilingual benchmarks and can be finetuned to outperform even larger models on specialized tasks.