The launch
On October 6, 2026, Google DeepMind released EmbeddingGemma 2, an open multimodal embedding model under Apache 2.0 (commercial use allowed). It turns text, code, images, video and audio, and combinations of them, into vectors in one shared 768-dimensional space. You can find a video clip from a voice memo, or search hours of recordings with a sentence.
The first EmbeddingGemma, text-only, passed 20 million downloads, per Google.
Specs
- 740 million parameters total, built on the Gemma 4 architecture.
- Modular: load only what you need. 270M for text and code, 440M with vision, 570M with audio, 740M for everything. All produce compatible vectors.
- 8,192-token context for text and code across 100+ languages.
- Matryoshka (MRL): vectors can be truncated to 512, 256 or 128 dimensions, for up to 6x less vector storage with little quality loss.
- 14% better than EmbeddingGemma 1 on MTEB (Code), keeping multilingual quality.
- Use via sentence-transformers (v6.1.0+):
SentenceTransformer("google/embeddinggemma-2"). Use bfloat16 or float32, not float16.
Why it matters
Embeddings underpin semantic search and RAG. Until now, companies with photos, inspection videos, call-center recordings and PDF manuals needed several models and separate pipelines. One small model running on a laptop, phone or CPU enables:
- Internal multimedia search: "the inspection video showing corrosion on the valve."
- On-device, offline RAG, without sending sensitive data to the cloud, which matters under GDPR, LGPD and Japan's APPI.
- Lighter pipelines: since it shares the tokenizer and audio encoder with Gemma 4, both run together with less memory.
Cautions
- Benchmark numbers are Google's; test on your own data, especially non-English.
- For video, tune frame sampling; long videos are expensive to embed.
- Truncating to 128 dimensions saves space, but Google mainly recommends it for text.
- Switching embedding models requires reindexing everything; plan the migration.
Getting started
- Pick a small use case (e.g., searching maintenance photos).
- Load only the encoders you need to save memory.
- Compare against your current embedding on accuracy and storage cost.
- Then expand to audio and video.
Sources
- Google Blog (2026-10-06): https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
- Google Developers Blog, "EmbeddingGemma 2: The Developer Guide": https://developers.googleblog.com/en/embeddinggemma-2-the-developer-guide/
- EmbeddingGemma 2 model card: https://ai.google.dev/gemma/docs/embeddinggemma/model_card_2
Want to build multimodal search and RAG on your own data? KGA IT offers a free consultation: kga-it.com.