EmbeddingGemma 2 adds multimodal local search
Google has released EmbeddingGemma 2 for local search across text, code, images, audio and video. The open model maps those inputs into a shared representation so that different media can be retrieved together.
Search can cross media without leaving a device
The Google release describes a 740 million parameter model built on Gemma 4 and distributed under Apache 2.0. Optional vision and audio encoders allow a smaller text-only configuration. Output vectors can also be shortened to reduce storage.
Find text and images locally
- Text-only model
- Optional vision and audio encoders omitted
- Multimodal encoders
- Retrieve across supported media
- Shorter output vectors
- Reduce storage
- Search result
- Retrieval capability, not a guaranteed answer
View data
| Configuration | Tradeoff |
|---|---|
| Text-only model | Optional vision and audio encoders omitted |
| Multimodal encoders | Retrieve across supported media |
| Shorter output vectors | Reduce storage |
| Search result | Retrieval capability, not a guaranteed answer |
Google · published 2026-10-06. Source-bound illustration, not a performance benchmark.
Download imageExamples include using a text query to locate an audio passage or a video moment. These are retrieval capabilities, not a claim that the model itself writes a reliable answer to every question. Google publishes evaluations, but local search quality still depends on the files, indexing and retrieval settings used.
Original sources
- Google: original announcementblog.google
Checked 9 Oct 2026 · A manually curated edition. Availability may change; company performance claims are not Trion test results. Editorial method.