Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- It is a multimodal embedding model: vectors help retrieve related content or match an intent, rather than directly generate a chat answer.
- The announcement offers Gallery image-search and video-moment demos, plus experimental Foresight for Mac. Device and store-region availability still depend on the relevant entry point.
- MediaPipe Tasks provides a more complete preprocessing and retrieval path. LiteRT and prequantized .litertlm bundles are another option when developers need finer control.
- Android ML Kit integration, NPU acceleration and automatic model updates are forward-looking in the October 6 original. Check subsequent release information before a production integration.
Start with retrieval, rather than generated answers
The Google AI Edge Team and Android ML Team introduced on-device uses for EmbeddingGemma 2 on October 6, 2026. The 740M-parameter open-weight model maps text, images, video frames and audio into one vector space for local media search, moment retrieval and intent matching.
A vector is a set of numbers that makes content easier to compare by meaning. A search system compares a query with indexed media and returns related items. It does not need to generate a text caption for every image first, and the embedding model does not thereby become a chat-answer generator.
A trial approach: start with samples, then non-sensitive media
Start from the announcement's Google AI Edge Gallery entry point, check which version is available for your device and begin with the supplied sample image set. The announcement links to Android and iOS downloads, but provides no complete compatibility list for every device, OS version or store region.
Next, compare text and image queries in Instant Media Search. Describe a scene for the first, or use an example image to find related content for the second. The announcement demonstrates rankings updating as a query is typed, as well as retrieval from a phone image or live camera view. Focus on whether the matches fit the intended search.
Then choose a short video of your own that contains no private material to explore Video Moments Finder. In a hypothetical sports clip, you could describe an action and inspect the matching timestamps after indexing. This example illustrates the relationship between a query and video segments; one match does not guarantee accurate recognition across every video.
Treat Foresight for Mac as an experimental app
The announcement also launches experimental Google AI Edge Foresight for Mac. It combines EmbeddingGemma 2 and Gemma 4 for local meeting-transcript, note and file retrieval, using relevant conversation details to enrich shorthand notes. This is a separate application experience, rather than another model-installation step inside Gallery.
Google describes these functions as locally processed and usable offline. When working with system audio, microphones, meetings or files, choose material according to the app's actual permissions and your organization's rules. Local execution does not itself grant permission to process any content, or establish that another integration follows the same data path.
Use MediaPipe for a prepared path, or LiteRT for more control
For developers, the announcement describes MediaPipe Universal Embedder and Semantic Retriever. The former handles media preprocessing and embedding generation; the latter handles local indexing and retrieval. It also describes Decision Task for matching inputs against candidate labels, such as intent routing. Check the documentation for the target platform before choosing an interface version.
For finer control over preprocessing, deployment and acceleration, the announcement points to LiteRT and prequantized .litertlm bundles. It describes running the same model file on CPU or GPU across supported edge platforms. A supported platform does not mean every device has the same performance or NPU support.
Read the ML Kit plan in its announcement context
The October 6 announcement describes Android ML Kit integration as coming in the following weeks. Its aim is a platform-integrated route that reduces model-management work for developers. That plan includes NPU acceleration when available and automatic model updates; this announcement alone does not establish that those features have reached every device.
The original gives no subsequent completion date or current interface version. It therefore also cannot establish that the integration remains unavailable on October 15. Before a production integration, check the target ML Kit version's release information, device support and model-management behavior, then choose the appropriate path.
Check memory, licensing and the scope of cost claims
The announcement gives a Google Pixel 11 Pro example: about 191 MB of active RAM for text-only weights and about 567 MB for the full multimodal model. It also describes loading only the modality encoders an application needs. These are figures for a particular device and configuration, rather than universal minimum requirements.
Open weights, software licensing, device costs and cloud-service charges are separate questions. The announcement mentions no cloud subscription costs specifically for the local Foresight experience. It provides no universal price or commercial-permission promise for every tool. Check the applicable license and download requirements for each chosen model bundle and component.
Sources
Frequently Asked Questions
Can my photos and videos be searched as soon as I have the model bundle?
The model bundle is not a completed media index. In the announced image workflow, embeddings are generated and stored in a local SQLite database; the video workflow first indexes visual and audio chunks. Initial preparation, supported formats and collection-size limits depend on the chosen app or implementation. A retrieval demo does not mean every file has already been processed.
Does local offline execution mean the first installation needs no network?
Separate obtaining the app, obtaining the model and running retrieval. The announcement describes relevant computations and app functions running offline on the device. It does not promise a disconnected first installation, download and update process on every platform. Prepare the required files, then confirm the offline scope of the target app's actual behavior.
Can I put the model straight into a commercial product?
The announcement calls the model open-weight, but does not set out complete commercial-license terms, redistribution obligations or the conditions for every dependency. Review the licenses and download requirements for the selected model bundle and components before commercial integration. Open weights and Foresight's lack of cloud subscription costs do not establish unrestricted or cost-free commercial use.