# Google’s EmbeddingGemma aims to raise the bar for on-device AI > Source: > Published: 2026-10-06 16:08:28+00:00 # Google’s EmbeddingGemma aims to raise the bar for on-device AI The new lightweight multimodal model is pitched as a multitasker that outperforms larger rivals, building on a predecessor that ran on under 200MB of RAM [Google](https://cryptobriefing.com/markets/alphabet/)’s Gemma family has a new entry that bets on small over big. EmbeddingGemma 2 is pitched as a lightweight multimodal model that handles multiple tasks and outperforms models larger than itself. ## What EmbeddingGemma 2 brings to the table The biggest change is the word “multimodal.” The original EmbeddingGemma was built as a text embedding model. Its successor is described as working across more than one type of input. An embedding model converts content into a list of numbers that captures its meaning. Those numerical fingerprints power search, recommendations, and the lookup step in retrieval-augmented generation, or RAG. In RAG, a chatbot fetches relevant documents before it answers, which helps keep it grounded. EmbeddingGemma 2 is described as handling multiple tasks rather than a single narrow job, and is said to beat larger models. ## The predecessor set a high bar Google DeepMind unveiled EmbeddingGemma on September 4, 2025, with a paper following around September 24–25, 2025. That first model packed 308 million parameters. Roughly 100 million were model parameters, and about 200 million were embedding parameters. ### AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. It drew on the Gemma 3 architecture, starting from a T5Gemma-style encoder-decoder adaptation. EmbeddingGemma consistently topped the Massive Text Embedding Benchmark (MTEB) leaderboards in Multilingual v2, English v2, and Code among open models under 500M parameters. It supported over 100 languages and offered a 2K token context window. Thanks to quantization-aware training, the model could run on less than 200MB of RAM. Inference clocked in at 15 to 22 milliseconds on EdgeTPU hardware. It also used Matryoshka Representation Learning, which lets developers choose output dimensions ranging from 768 down to 128. The original was distributed through Hugging Face, Ollama, Kaggle, and Google under a responsible commercial license. ## Why on-device efficiency matters When embeddings are generated on the device, the underlying data does not have to leave it. The first EmbeddingGemma was explicitly aimed at privacy-sensitive applications. Offline RAG, semantic search, classification, and clustering on phones and laptops were all core use cases for the original release. **Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).