EmbeddingGemma 2: One On-Device Embedder for Five Modalities

Google DeepMind released EmbeddingGemma 2 on 6 October 2026. The first generation was text only and passed 20 million downloads; this one pulls code, images, video and audio into a single 768-dimension space at 740M parameters under Apache 2.0, running in about 191MB of quantised memory for text on a phone

Tool Interface

Interactive tool will be available soon

Features

✓ One model, five inputs: text, code, images, video frames and audio share a vector space, so cross-modal similarity works too
✓ Modular loading: 270M for text, 440M with vision, 570M with audio, 740M for everything, all projecting into the same space
✓ Truncatable vectors: Matryoshka representation learning cuts 768 dimensions to 512, 256 or 128 for up to 6x less local vector storage
✓ A big jump on code: MTEB Code rises from 68.76 to 78.68, a 9.92-point gain that suits local codebase indexing and semantic search
✓ Built for the edge: an 8K context window, four times its predecessor, handles about 5.5 minutes of audio, 29 images or 58 video frames, with full multimodal weights at about 567MB on a Pixel 11 Pro

How to Use

Grab the weights from Hugging Face or Kaggle under Apache 2.0, with Ollama, MLX, vLLM, llama.cpp and sentence-transformers support
Load only what you need: a text-only retriever should start at 270M rather than pulling the full multimodal weights
Decide dimensions before you build the index: 768 is most accurate, and truncating to 512 or 256 is far cheaper now than migrating later
Measure recall on your own corpus, especially for code and audio, instead of trusting the paper scores alone

FAQ

When was EmbeddingGemma 2 released?

6 October 2026. Official announcement: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2 , model card: https://huggingface.co/google/embeddinggemma-2 .

Is it 740M or 744M parameters?

Google's own blog and the Hugging Face model card both state 740M total parameters (270M text, 170M vision, 300M audio). Some secondary write-ups say 744M; trust the official figure.

Does it actually run on a phone?

Google's figures are about 191MB of active memory for text-only weights and roughly 567MB for full multimodal, both quantised, on a Pixel 11 Pro. Mid-to-high-end phones can handle it, but the full multimodal build is about three times heavier than text-only.

Can I use it commercially?

The model is Apache 2.0 licensed and Google describes the licence as commercially permissive. Still verify the model card's terms yourself before shipping.

How does it relate to Gemini Embedding?

Google says EmbeddingGemma 2 is built with the same technology as the Gemini Embedding models, but aimed at on-device and privacy-first use without a cloud API. For large hosted workloads, the Gemini Embedding service remains the option.