Google Releases EmbeddingGemma 2 for On-Device Multimodal Use: what are the key facts?

Google released the EmbeddingGemma 2 embedding model on October 7.

Google Releases EmbeddingGemma 2 for On-Device Multimodal Use: what are the key facts?

The model is built on the Gemma 4 architecture and is licensed under Apache 2.0.

Google Releases EmbeddingGemma 2 for On-Device Multimodal Use: what are the key facts?

It has 740 million total parameters and supports text, code, images, video, and audio.

Google Releases EmbeddingGemma 2 for On-Device Multimodal Use: what are the key facts?

After quantization, the text-only weights require only about 191MB of memory on a Pixel 11 Pro.

Google Releases EmbeddingGemma 2 for On-Device Multimodal Use: what are the key facts?

Its MTEB Code score rose from 68.76 to 78.68, a gain of 9.92 points.

Physix Frontier · News Briefing Card (enbrief · Oct 7, 2026)

Google Releases EmbeddingGemma 2 for On-Device Multimodal Use

KEY FACTS

Google released the EmbeddingGemma 2 embedding model on October 7.
The model is built on the Gemma 4 architecture and is licensed under Apache 2.0.
It has 740 million total parameters and supports text, code, images, video, and audio.
After quantization, the text-only weights require only about 191MB of memory on a Pixel 11 Pro.
Its MTEB Code score rose from 68.76 to 78.68, a gain of 9.92 points.

KEY DATA

740MTotal Parameters

191MBText-Only Runtime Memory

567MBFull Multimodal Runtime Memory

78.68MTEB Code Score

PHYSIX OBSERVATION

EmbeddingGemma 2 squeezes multimodal embeddings into phone memory, so offline cross-modal retrieval is no longer exclusive to the big players. Google is courting developers with Apache 2.0 and an on-device toolchain, and the signal is clear: embedding models are shifting from cloud APIs to on-device infrastructure, and their privacy and latency advantages will spawn a wave of local RAG applications.

Source: enbrief original report ↗