Welcome EmbeddingGemma, Google's new efficient embedding model (huggingface.co)

Google has introduced EmbeddingGemma, a compact and efficient multilingual embedding model optimized for on-device applications. With just 308 million parameters and a 2,048-token context window, EmbeddingGemma supports over 100 languages and achieves state-of-the-art performance on the Massive Text Embedding Benchmark (MTEB) while requiring under 200MB of RAM when quantized. This model is particularly significant for the AI/ML community as it balances high accuracy and multilingual versatility with low resource demands, enabling enhanced semantic search, retrieval-augmented generation, and other natural language processing tasks on mobile and edge devices. Technically, EmbeddingGemma builds on Google DeepMind’s Gemma3 transformer backbone, redesigned with bi-directional attention to function as an encoder rather than a decoder, which improves performance on embedding tasks like retrieval. It outputs 768-dimensional text embeddings derived from token embeddings via mean pooling and dense transformation layers. The model is trained using Matryoshka Representation Learning (MRL), allowing users to truncate embeddings to 512, 256, or 128 dimensions for faster, more memory-efficient downstream processing without significant loss in quality. Trained on a massive, carefully curated dataset of 320 billion tokens, EmbeddingGemma has demonstrated superior performance versus larger models on benchmarks and specialized tasks, such as medical passage retrieval after finetuning. EmbeddingGemma integrates seamlessly with popular frameworks like Sentence Transformers, LangChain, and LlamaIndex, with built-in support for prompt-based task conditioning to optimize embeddings for various applications, from classification to semantic similarity and code retrieval. Its open-source availability and practical design make it a powerful new tool for developers seeking multilingual, high-quality embeddings that run efficiently on everyday hardware.