Google DeepMind Launches EmbeddingGemma 2 Multimodal On-Device Embedding Model

Google DeepMind released EmbeddingGemma 2, a 740M-parameter multimodal embedding model that unifies text, code, images, audio, and video in a single shared embedding space, designed to run fully on-device. The model is built on the Gemma 4 architecture and released under an Apache 2.0 license. It is modular: a 270M-parameter text-only base with optional 170M vision and 300M audio encoders that load on demand. On a Google Pixel 11 Pro, quantized text weights require approximately 191MB RAM, and the full multimodal configuration uses approximately 567MB. The context window is 8K tokens, supporting roughly 5.5 minutes of audio, 29 images, or 58 video frames in a single input. Using Matryoshka Representation Learning, output vectors can be truncated from 768 to 128 dimensions, enabling up to 6x storage reduction for local vector databases. Google reports a 9.92-point improvement over the original EmbeddingGemma on the MTEB Code benchmark (68.76 to 78.68), and claims top sub-1B results on audio and vision benchmarks. The predecessor EmbeddingGemma surpassed 20 million downloads; the new model is available on Hugging Face and Kaggle, with support from tools including LitERT, Google AI Edge, MediaPipe, Qdrant, Unsloth, Ollama, LMStudio, vLLM, and llama.cpp.