Product

EmbeddingGemma 2

EmbeddingGemma 2 is an open multimodal vector model from Google that represents text, code, images, video, and audio in a single 768-dimensional space. The model converts various media types into numerical vectors, allowing developers to compare different content formats directly for search and retrieval applications. Currently, EmbeddingGemma 2 is available in two distinct configurations. The text and code version features 270 million parameters, while the full multimodal version contains 740 million parameters. According to Google, the 740-million-parameter model outperforms competitors twice its size on multimodal benchmarks. The model runs locally on devices via WebGPU, requiring about 191 MB of RAM and no API key. This setup supports offline retrieval-augmented generation and can reduce local vector database storage requirements by up to six times. When configured to shorter 128-dimensional vectors, EmbeddingGemma 2 retains around 90 percent of its text retrieval quality and approximately 75 percent of its performance for image, video, and speech retrieval. Recent developments show its integration into consumer software, such as Google AI Edge Foresight, a local-first meeting note-taker for Mac that utilizes EmbeddingGemma 2 to function offline without an internet connection.

5 stories · Updated

Key facts

  • Google released EmbeddingGemma 2 as an open model that represents text, code, images, video, and audio in a single 768-dimensional space.
  • The full multimodal version of EmbeddingGemma 2 has 740 million parameters, while the text and code configuration has 270 million parameters.
  • EmbeddingGemma 2 runs locally via WebGPU, uses about 191 MB of RAM, and requires no API key.
  • At 128 dimensions, EmbeddingGemma 2 retains about 90% of text retrieval quality and around 75% for image, video, and speech retrieval.
  • The local-first Mac application Google AI Edge Foresight uses EmbeddingGemma 2 to operate without an internet connection.

Timeline

October 2026