Google DeepMind Releases Open Multimodal EmbeddingGemma 2 Model
Google DeepMind introduced the open EmbeddingGemma 2 model for creating shared vector representations of text, code, images, video, and audio.
On October 6, 2026, Google DeepMind announced the EmbeddingGemma 2 model, an open multimodal model designed to convert different types of content into a shared vector space. It works with text, including source code, images, video, and audio, producing 768-dimensional embeddings.
Embeddings are numerical representations of content that applications use, for example, in semantic search, classification, or retrieval systems for RAG. The shared space is intended to enable comparisons across supported modalities.
EmbeddingGemma 2 and its parameters
The model has a total of 740 million parameters. The text component contains 270 million parameters. The image encoder is optional and has 170 million parameters, while the audio encoder is also optional and has 300 million parameters. Deployments that work only with text therefore do not need to use the optional image and audio components.
The EmbeddingGemma 2 model supports a context window of 8,192 tokens. Developers can shorten the 768-dimensional embeddings to 512, 256, or 128 dimensions. This reduction may be relevant where an application needs to reduce the storage or processing requirements for vectors.
Open weights and local use
Google released the model under the Apache 2.0 license. The weights are publicly available in Google’s repository on the Hugging Face platform. The license facilitates integration into custom applications and projects that want to run the model locally.
According to the announcement, the model is also intended for local semantic search, classification, and RAG scenarios involving a combination of text, code, images, video, and audio. Local processing may reduce latency in such cases and limit the need to send sensitive content to a remote server. However, actual memory requirements and performance will depend on the quantization used, the hardware, the library, and the specific task.
What will need to be verified in practice
Google published its own benchmarks, but claims about the model’s position among similarly sized multimodal embedding models do not yet have an independently verified comparison. Reproductions of the results and tests against other models with fewer than one billion parameters will be important.
Developers will also monitor compatibility and real-world performance in Transformers, sentence-transformers, MediaPipe, and LiteRT. Open questions also include practical limits in multilingual tasks, video and audio processing, and how to address bias, privacy, and malicious-use risks in applications built on embeddings.
Sources
- Google Blog / Google DeepMind – Confirms the announcement date, model release, supported modalities, parameter count, license, context window, and stated local deployment scenarios.
- Google AI for Developers – EmbeddingGemma 2 model card – Lists the technical parameters, benchmark results, supported embedding dimensions, input limits, and known model limitations.
- Hugging Face – google/embeddinggemma-2 – Confirms the model’s public availability from the Google account and the Apache 2.0 license.
Verified and updated: 10/07/2026 06:24
