Google DeepMind releases EmbeddingGemma 2, a lightweight multimodal embedding model optimized for on-device inference

Google DeepMind introduced EmbeddingGemma 2, a 740M-parameter multimodal embedding model designed for on-device inference and released under an Apache 2.0 license. The model natively maps combinations of text, images, audio, and video into a unified embedding space, enabling cross-modal semantic search and retrieval directly on consumer hardware.

The model achieves leading scores among sub-1B multimodal embedders across benchmarks such as the Massive Text Embedding Benchmark (MTEB), MTEB Code, and the Massive Audio Embedding Benchmark (MAEB). It delivers a 9.92-point improvement on code performance in MTEB Code, increasing from 68.76 to 78.68, and sets a new standard in quality-per-parameter for sub-1B models across image, video, document, and audio tasks.

EmbeddingGemma 2 is built on the Gemma 4 architecture and supports modular configurations: as little as 270M parameters for text-only workloads, with optional vision (170M) and audio (300M) encoders for full multimodal support. It uses Matryoshka Representation Learning (MRL) to dynamically truncate output vectors from 768 dimensions down to 512, 256, or 128 dimensions, enabling up to 6x storage reduction for local vector databases and memory usage.

The model is optimized for on-device performance, running efficiently within tight resource constraints. On a Google Pixel 11 Pro, EmbeddingGemma 2 requires approximately 191MB of active RAM for text-only weights and approximately 567MB for the full multimodal model when using quantization. It features an 8K token context window, allowing it to process up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations directly on local hardware.

EmbeddingGemma 2 is positioned for privacy-first retrieval augmented generation (RAG) pipelines and cross-modal search applications. It can power use cases such as finding a specific video clip from a voice memo or searching through hours of audio recordings based on a text query, all processed by a single, natively multimodal model. The model shares its text tokenizer and audio encoder with Gemma 4, enabling unified pipelines with a lower combined total memory footprint.

Developers can experiment with EmbeddingGemma 2 in the Google AI Edge Gallery through tools like Instant Media Search, Video Moments Finder, and the Google AI Edge Foresight app, which pair the model with Gemma 4 for contextual reasoning and real-time decision engines.