Google DeepMind introduced EmbeddingGemma 2, an Apache 2.0 model that creates shared embeddings for text, code, images, video, and audio.
Google DeepMind announced EmbeddingGemma 2 on October 6, 2026, expanding its EmbeddingGemma line from text embeddings to a model that can represent text, code, images, video, and audio in a shared embedding space. The model has 740 million parameters, is built on the Gemma 4 architecture, and is released under the Apache 2.0 license.
The company says the model is designed for on-device use, including search across different media types. Its 8K-token context window is four times larger than EmbeddingGemma 1’s, and Google says it can process up to 5.5 minutes of audio, 29 images, or 58 video frames, including interleaved combinations. With quantization on a Google Pixel 11 Pro, the reported active memory requirement is about 191MB for text-only weights and 567MB for the full multimodal model.
Developers can also reduce embedding-vector size using Matryoshka Representation Learning, truncating outputs from 768 dimensions to 512, 256, or 128. Google says this can cut local vector-database storage and memory use by up to six times. For text-only workloads, the modular configuration can use as little as 270 million parameters, with optional vision and audio encoders for multimodal capabilities.
Google reports a 9.92-point gain over the first EmbeddingGemma on the MTEB Code benchmark, increasing from 68.76 to 78.68. It also says the model achieves leading results among sub-billion-parameter multimodal embedding models on benchmarks including MTEB Code and MAEB, while matching or exceeding some larger models on other tasks.
Model weights are available through Hugging Face and Kaggle. Google says availability in Gemini Enterprise Agent Platform Model Garden is coming soon. The company positions local embedding generation as a way to support offline retrieval and reduce reliance on sending data to remote systems.