Google’s EmbeddingGemma 2 Makes Multimodal Retrieval Portable

What changed · following since Oct 7 Oct 15 — EmbeddingGemma 2 was adapted into an on-device speech-to-text system that transcribes an hour of audio in 12.6 seconds on a base M3 MacBook Air.

EmbeddingGemma 2 combines portability with a broader multimodal surface than a conventional text-only embedding service. Its open weights and shared representation can support self-hosted retrieval while leaving room for specialized local applications built around the same base model.

• Google describes EmbeddingGemma 2 as an open, lightweight multimodal embedding model optimized for privacy-first use cases.
• The model is available under the Apache 2.0 license, according to Simon Willison’s October 6 discussion.
• Embedding systems commonly generate thousands or millions of vectors for later comparison, so changing the embedding model can create a substantial re-embedding workload.
• The open model can be self-hosted if a hosted provider later stops offering the service, while teams can still choose to pay a provider for inference.
• An adaptation maps short audio slices into EmbeddingGemma 2’s text space and selects among approximately 229,000 English words to produce transcripts.
• Caveat: the adaptation trains only a small reader; the base model was never trained for speech-to-text.

Bottom line EmbeddingGemma 2 is worth evaluating not only as a portable retrieval model but also as a foundation for local multimodal tooling.