EmbeddingGemma
Google's compact open embedding model, a miniaturised relative of Gemini, that turns text into meaning-carrying vectors for search understanding.
Google has released an open, multilingual embedding model called EmbeddingGemma. It is essentially a miniature relative of Gemini, the system behind Google's advanced search. Because Gemma is the open-source sibling of Gemini, it gives us a rare public look into the kind of technology Google uses to power its search infrastructure.
EmbeddingGemma is built for practical, everyday use. With around three hundred million parameters, it is small enough to run directly on your device. It supports over one hundred languages and offers top-tier quality for its size. Unlike models built to generate text, this one uses bidirectional attention, meaning it is specifically optimized for deep understanding.
One of its most impressive features is something called Matryoshka Representation Learning. This allows you to shrink the model's output vector from its standard size down to much smaller dimensions with very little loss in quality. It gives you the flexibility to trade a tiny bit of accuracy for much faster speeds and lower storage requirements on demand. This versatility is why these embeddings are so useful for complex tasks, including expanding and refining search queries.
EmbeddingGemma is Google's open, multilingual embedding model — effectively a miniaturised relative of Gemini, the system behind Google's advanced search. Because Gemma is the open "little sister" of Gemini, EmbeddingGemma is a rare public window into the kind of representations Google's search infrastructure relies on.
Its specs are built for practical use: about 308M parameters (small enough to run on-device), a 2K-token context window, 768-dimensional output vectors, and support for 100+ languages, with state-of-the-art quality for its size. It uses encoder-style bidirectional attention, optimised for understanding rather than generation.
A standout feature is Matryoshka Representation Learning, which lets the 768-dim output be truncated to 512, 256 or 128 dimensions with little quality loss — trading accuracy for speed and storage on demand. We use Gemma-family embeddings across our own work, including query fan-out.