EmbeddingGemma 2 multimodal embedding model

Google DeepMind released EmbeddingGemma 2 on October 6, 2026, describing it as its first natively multimodal open embedding model designed for on-device use. The model has 740M parameters built on the Gemma 4 architecture and handles text, code, images, audio, and video in a unified embedding space. Its modular design allows a text-only footprint of 270M parameters, with vision (170M) and audio (300M) encoders loaded on demand. On a Pixel 11 Pro, text-only operation uses approximately 191MB RAM and the full multimodal stack uses approximately 567MB. The 8K token context window is four times larger than EmbeddingGemma 1, and supports roughly 5.5 minutes of audio, 29 images, or 58 video frames in a single input. Google reports code search performance improved from an MTEB Code score of 68.76 to 78.68 over the predecessor. Using Matryoshka Representation Learning, developers can truncate output vectors from 768 to 128 dimensions for up to 6x storage reduction in local vector databases. The model is released under Apache 2.0 and is available through Hugging Face, Kaggle, and a range of inference tools. The model can run directly on consumer devices including laptops, phones, and browsers, keeping searches private without sending data to the cloud.