Google launches EmbeddingGemma 2, 740 million-parameter Gemma 4-based on-device model
Google just rolled out EmbeddingGemma 2, a major upgrade to its open-source model that turns stuff like text and audio into numbers your device can process quickly.
With 740 million parameters and built on Gemma 4 architecture, it's designed to make apps smarter right on your phone or laptop (no cloud needed).
EmbeddingGemma 2 supports 270 million-740 million parameters
You can tweak EmbeddingGemma 2 for lighter or heavier tasks (from 270 million up to 740 million parameters), so it works for everything from coding help to handling multiple data types at once.
It uses Matryoshka Representation Learning to save space while retaining much of the original quality.
Plus, since it runs locally and is open under the Apache 2.0 license, it's great for privacy-focused apps.
As Google CEO Sundar Pichai puts it, this model is "Introducing EmbeddingGemma 2, a new open multimodal model that sets the standard for on-device efficiency."