Google releases EmbeddingGemma 2 for on-device multimodal search

The company also positions EmbeddingGemma 2 as the retrieval half of an on-device retrieval-augmented generation setup. In that pairing, EmbeddingGemma 2 finds relevant local files, while Gemma 4 reasons over the retrieved material to produce an answer. Google says this keeps the process local and reduces the memory footprint.

A modular footprint

EmbeddingGemma 2 can be configured from 270 million to 740 million parameters, according to Google DeepMind developer-experience lead Omar Sanseviero. He also says it uses Matryoshka embeddings for storage efficiency.

At its largest size, Google says the model uses about 191MB to 567MB of active RAM. The company also reports an 8K context window, four times larger than the first generation, with capacity for as much as 5.5 minutes of audio, 29 images or 58 video frames in a single pass.

Google claims the 740-million-parameter version outperforms some specialist models more than twice its size. Sundar Pichai said the model’s weights were available on Hugging Face when Google announced the release on October 6.

Many accounts welcomed EmbeddingGemma 2 for its strong on-device multimodal performance that beats larger models, while some replies complained about other Gemini versions or rival embedding tools.

Based on 189 sentiment-bearing replies from 123 accounts across 2 conversations.

Many accounts welcomed EmbeddingGemma 2 for its strong on-device multimodal performance that beats larger models, while some replies complained about other Gemini versions or rival embedding tools.

Based on 189 sentiment-bearing replies from 123 accounts across 2 conversations.