The architecture features an 8K context window, which is four times larger than the original text only version, to optimize on device efficiency for mobile and desktop. When paired with Gemma 4, the model powers privacy first retrieval augmented generation by accessing and reasoning over local files entirely offline. Support for the Gemini Enterprise Agent Platform Model Garden follows.
Tuesday, October 20, 2026
Story
Google Launches EmbeddingGemma 2 as First Open Multimodal Embedding Model
AI14 days ago
Up to 58 video frames or 5.5 minutes of audio can be processed in a single pass to enable local search and retrieval. Google and Google DeepMind released EmbeddingGemma 2, a 740M parameter open model that unifies text, code, image, video, and audio into a single embedding space under an Apache 2.0 license. The system outperforms some specialist models over twice its size, utilizes between 191MB and 567MB of active RAM, and provides a 9.92 point improvement in code performance over the previous version. Weights are available now on Hugging Face and Kaggle.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.