Generative AI & Models6 October 2026, 19:57 UTC3 min read

EmbeddingGemma 2: an open, lightweight multimodal embedding model

The model builds on the Gemma 4 architecture, contains 740 million parameters and is released under an Apache 2.0 license. It follows the original EmbeddingGemma, which saw more than 20 million downloads for on‑device search and privacy‑first RAG pipelines.

1 source primary source

Key points

Model achieves leading scores among sub‑1 B multimodal embedders and improves code benchmark by 9.92 points
8 K token context window enables processing of up to 5.5 min audio, 29 images or 58 video frames locally

EmbeddingGemma 2 achieves top scores among sub‑1 billion‑parameter multimodal embedders on benchmarks such as MTEB Code and MAEB, improving its code benchmark by 9.92 points (from 68.76 to 78.68). It offers an 8 K token context window—four times larger than its predecessor—allowing up to 5.5 minutes of audio, 29 images or 58 video frames to be processed locally. With Matryoshka Representation Learning, output vectors can be truncated to 512, 256 or 128 dimensions, delivering up to six‑fold storage reduction.

Developers can download the weights from Hugging Face or Kaggle, deploy on‑device via Google AI Edge MediaPipe or LiteRT, and fine‑tune using tools like Unsloth. The model aims to enable private, low‑latency, cross‑modal search and retrieval entirely on edge hardware.

Full story from Google DeepMind · by Sahil Dua primary sourceOpen source ↗