EmbeddingGemma 2 is the most capable on‑device multimodal embedding model, natively mapping text, images, audio and video into a shared vector space. Built on the Gemma 4 architecture, it has 740 M parameters and is released under the permissive Apache 2.0 license, making it suitable for consumer‑grade inference.
On benchmarks such as MTEB and MAEB, EmbeddingGemma 2 achieves leading scores among sub‑1B multimodal embedders and even outperforms larger specialist models on many tasks. Its modular design allows a text‑only configuration with 270 M parameters, while optional vision (170 M) and audio (300 M) encoders can be added as needed.
Matryoshka Representation Learning (MRL) lets developers truncate the output from 768 dimensions down to 512, 256 or 128, cutting storage requirements by up to six times. With quantization, the model runs on a Google Pixel 11 Pro using roughly 191 MB RAM for text‑only weights and about 567 MB for the full multimodal stack.
An 8K token context window translates to roughly 5.5 minutes of audio, 29 images or 58 video frames, enabling a single pass over mixed‑modal inputs on local hardware. Code performance improves by 9.92 points (from 68.76 to 78.68 on MTEB Code), making it ideal for on‑device codebase indexing and semantic code search.
Local embedding generation safeguards privacy, reduces latency, and can be paired with generative models like Gemma 4 for fully offline retrieval‑augmented generation pipelines. The Google AI Edge Gallery showcases instant media search, video moment finding and cross‑modal retrieval demos.
Getting started options include:
- Downloading weights from Hugging Face or Kaggle;
- Deploying with Google AI Edge MediaPipe, LiteRT, or in browsers via transformers.js/WebGPU;
- Serving the model with transformers, sentence‑transformers, MLX, vLLM, llama.cpp, SGLang, Ollama or LMStudio;
- Storing vectors in Qdrant;
- Fine‑tuning following Unsloth’s guide.
Comprehensive documentation, example code and inference guides are available to accelerate on‑device semantic search, routing and retrieval.
Review