EmbeddingGemma 2: an open, lightweight multimodal embedding model
Google releases EmbeddingGemma 2, an open-source model that maps text, images, audio and video into one shared embedding space so they can be searched and compared together. Built on the Gemma 4 architecture with 740 million parameters, it is designed for local devices and supports shrinking its output vectors.
What the article says
On October 6, 2026, Google introduces EmbeddingGemma 2, an open-source, lightweight, multimodal embedding model that places text, images, audio and video in a single vector space. Built on Gemma 4 with 740 million parameters and released under the Apache 2.0 license, it follows a first EmbeddingGemma that passed 20 million downloads. It is modular: 270 million parameters are enough for text only, with optional vision (170M) and audio (300M) encoders. Its context window reaches 8,000 tokens, four times that of version 1.
Google reports the best scores among multimodal models under one billion parameters on MTEB Code and MAEB, with MTEB Code rising from 68.76 to 78.68. The 768-dimension vectors can be truncated down to 128 to cut storage. The model targets on-device use, with about 191 MB of RAM for text only on a Pixel 11 Pro. Weights are on Hugging Face and Kaggle, with many compatible tools.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Google DeepMind. The content and rights belong to the publisher: please read the original article.