Blog
Google releases EmbeddingGemma 2, an open on-device model for searching text, images, audio and video
On 6 October 2026 Google DeepMind released EmbeddingGemma 2, a 740M-parameter open model under Apache 2.0. It maps text, code, images, audio and video into one 768-dimensional space, so apps can search across media types on the device.
On 6 October 2026 Google DeepMind released EmbeddingGemma 2, an open embedding model that handles text, code, images, audio and video in one model. It is small enough to run on phones and laptops, and it is licensed under Apache 2.0, which allows commercial use.
What an embedding model does
An embedding model turns content into a list of numbers (a vector) so software can measure how similar two things are. It powers semantic search, recommendations, classification, and retrieval-augmented generation (RAG), where an AI looks up relevant material before it answers.
The first EmbeddingGemma handled text only. Google says it has been downloaded more than 20 million times.
What is new in version 2
According to Google's announcement and developer guide:
- One shared space for every type of media. Text, code, images, video and audio all map into the same 768-dimensional vector space. So a text query can find a video clip, or a voice memo can find a related image, without first converting everything to text.
- 740 million parameters for the full multimodal model, built on the Gemma 4 architecture.
- Modular loading. You can load only the parts you need, from about 270M parameters for text and code up to 740M for all media types.
- Runs offline on consumer CPUs and GPUs, according to Google's documentation.
On-device numbers
Google's AI Edge team says that on a Pixel 11 Pro the text-only weights need about 191MB of active RAM, and the full multimodal model about 567MB. The team argues that one model can replace a chain of separate captioning, speech-to-text and text-embedding models, cutting latency and memory use.
Google also highlights zero-shot intent routing: matching a user's input against a list of labels or descriptions in milliseconds, with no training data or fine-tuning. That suits apps that need to decide quickly which feature or tool should handle a request.
Where to try it
- Google AI Edge Gallery on mobile and Google AI Edge Foresight on Mac show interactive demos.
- Google says that in the coming weeks the model will also be offered as a service on Android through ML Kit, with NPU acceleration. That part is not available yet.
How to read the claims
Google calls EmbeddingGemma 2 "best-in-class" for its size. That is Google's own judgement based on its own evaluations. As with any embedding model, the real test is how well it retrieves results from your own data, so benchmark it on a sample before switching.
What this means for you
For developers building search or RAG features, especially private, offline ones on phones or laptops, this is a practical release. You get one open model, a commercial-friendly licence, and a single index that can cover documents, photos, recordings and video. For a small business, it makes local search across catalogues, product photos and voice notes more realistic without sending data to a cloud API.