Google DeepMind has released EmbeddingGemma 2, an open-weight embedding model that converts text, images, audio, video and code into numerical vectors in a single shared space, so apps can search across media types without sending data to the cloud. The launch was announced on October 6, 2026, on Google's official blog.
What it does
EmbeddingGemma 2 is built on the Gemma 4 architecture and released under the commercially permissive Apache 2.0 license. With 740 million parameters, it is designed for on-device inference: Google says it can find a specific video clip from a voice memo, or pull material from hours of audio recordings using a text query. Because embeddings are generated locally, the company says the model supports data privacy, lower latency and fully offline cross-modal search and retrieval.
Modular and compact
The model is modular by design. Text-only workloads need as little as 270 million parameters, with optional vision (170M) and audio (300M) encoders for full multimodal support. Google says that with quantization, the text-only version uses about 191 MB of active RAM on a Pixel 11 Pro, while the full multimodal model uses roughly 567 MB. Using Matryoshka Representation Learning, developers can truncate output vectors from 768 dimensions down to 128, cutting local vector-database storage by up to six times.
Benchmarks and availability
Google claims best-in-class scores among sub-1B multimodal embedders, with a 9.92-point jump on MTEB Code (from 68.76 to 78.68), making it suitable for local codebase indexing and coding-agent retrieval. It also outperforms some specialist models more than twice its size on vision and audio tasks. Context is 8K tokens, enough for up to 5.5 minutes of audio, 29 images or 58 video frames.
Weights are available on Hugging Face and Kaggle, with support coming to Google's ML Kit for Android in the coming weeks, plus MediaPipe tooling for iOS, macOS, Windows, Linux and the web.
Sources: Google Blog: EmbeddingGemma 2 is a best-in-class open model for natively multimodal embeddings, The Decoder: Google claims EmbeddingGemma 2 outperforms rival embedding models twice its size
What it does
EmbeddingGemma 2 is built on the Gemma 4 architecture and released under the commercially permissive Apache 2.0 license. With 740 million parameters, it is designed for on-device inference: Google says it can find a specific video clip from a voice memo, or pull material from hours of audio recordings using a text query. Because embeddings are generated locally, the company says the model supports data privacy, lower latency and fully offline cross-modal search and retrieval.
Modular and compact
The model is modular by design. Text-only workloads need as little as 270 million parameters, with optional vision (170M) and audio (300M) encoders for full multimodal support. Google says that with quantization, the text-only version uses about 191 MB of active RAM on a Pixel 11 Pro, while the full multimodal model uses roughly 567 MB. Using Matryoshka Representation Learning, developers can truncate output vectors from 768 dimensions down to 128, cutting local vector-database storage by up to six times.
Benchmarks and availability
Google claims best-in-class scores among sub-1B multimodal embedders, with a 9.92-point jump on MTEB Code (from 68.76 to 78.68), making it suitable for local codebase indexing and coding-agent retrieval. It also outperforms some specialist models more than twice its size on vision and audio tasks. Context is 8K tokens, enough for up to 5.5 minutes of audio, 29 images or 58 video frames.
Weights are available on Hugging Face and Kaggle, with support coming to Google's ML Kit for Android in the coming weeks, plus MediaPipe tooling for iOS, macOS, Windows, Linux and the web.
Sources: Google Blog: EmbeddingGemma 2 is a best-in-class open model for natively multimodal embeddings, The Decoder: Google claims EmbeddingGemma 2 outperforms rival embedding models twice its size
