EmbeddingGemma 2: Features, Specs and How to Use It

What is EmbeddingGemma 2?

EmbeddingGemma 2 is an open-source, lightweight multimodal embedding model developed by Google DeepMind. Released on 6 October 2026, it maps combinations of text, images, audio, video, and code into a single, unified embedding space.

This model builds on the success of the original text-only model, which received more than 20 million downloads. It is designed to run efficiently on consumer hardware, enabling developers to build fast, offline search engines and privacy-focused AI applications.

What are the key features of EmbeddingGemma 2?

EmbeddingGemma 2 delivers powerful multimodal capabilities while remaining small enough to run entirely on local devices. It introduces several major efficiency improvements over the first version of the model.

The model is modular by design, meaning you do not have to load all encoders at once. It requires 270 million parameters for text-only workloads, with optional vision (170 million) and audio (300 million) encoders.

Using Matryoshka Representation Learning, developers can dynamically truncate output vectors from 768 dimensions down to 512, 256, or 128 dimensions. This flexibility provides up to a 6x reduction in local database storage and active memory usage.

The extended context window allows the model to process up to 5.5 minutes of audio, 29 images, or 58 video frames directly on local hardware. When quantized, it runs smoothly on mobile hardware like the Google Pixel 11 Pro.

How does EmbeddingGemma 2 perform on benchmarks?

EmbeddingGemma 2 achieves top-tier scores among sub-1B parameter models across multiple language, vision, and audio tasks. According to Google DeepMind, the model matches or outperforms several larger specialist models.

The model shows a major improvement in processing programming languages. It delivers a 9.92-point increase on the Massive Text Embedding Benchmark (MTEB) Code dataset compared to its predecessor.

What are the best use cases for EmbeddingGemma 2?

EmbeddingGemma 2 is built specifically for on-device indexing, search, and retrieval pipelines. Processing embeddings locally allows developers to maintain strict data privacy and eliminate network latency.

Because it shares a text tokenizer and audio encoder with Gemma 4, developers can run both models together in a unified pipeline. This combination allows you to build local retrieval augmented generation (RAG) pipelines with a lower memory footprint.

How to get and use EmbeddingGemma 2?

Developers can easily access EmbeddingGemma 2 weights to build local cross-platform applications. Google DeepMind has worked with ecosystem partners to ensure immediate compatibility with popular AI frameworks.

1. Download the model weights from Hugging Face or Kaggle. For models optimized specifically for local hardware, visit the LiteRT Community on Hugging Face.
2. Choose an on-device deployment framework. Use Google AI Edge MediaPipe for quick embedding, retrieval, and decision tasks, or use LiteRT for custom model integration.
3. Serve the model locally using your preferred development tools, such as transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, or LMStudio.
4. Store your generated embedding vectors using a local vector database like Qdrant.
5. If you need to customize the model for specific business data, follow fine-tuning guides provided by Unsloth.

Frequently asked questions

Sources: Google · Google