Google's new open embedding model can search your voice memos and videos on-device, without uploading anything

Google DeepMind ships a small multimodal embedding model that works offline on your device

Google DeepMind announced EmbeddingGemma 2 on October 6, 2026, releasing an open-weight embedding model designed to run directly on phones and laptops and, for the first time in this product line, to handle images, video and audio alongside text. The weights are published on Hugging Face and Kaggle under an Apache 2.0 license, which permits commercial use. The announcement appeared on the Google DeepMind blog under the title "EmbeddingGemma 2: an open, lightweight multimodal embedding model", credited to research engineers Sahil Dua and Henrique Schechter Vera.

The practical meaning for a non-specialist is straightforward: this is a small, free, openly licensed model that can search your own photos, voice memos and videos with a text question, entirely on your device. Nothing has to be uploaded to a server, the search works with no internet connection, and small developers can build on it without paying a cloud provider per query.

Google, as the vendor reporting on its own model, states that EmbeddingGemma 2 has 740 million parameters, runs within roughly 191MB of active RAM for text-only workloads and roughly 567MB for the full multimodal model on a Pixel 11 Pro with quantization, and achieves leading benchmark scores among sub-1B multimodal embedders on MTEB Code and MAEB. Those figures are Google's own claims and have not been independently verified for this article. The weights themselves, the license and the model's existence are independently confirmed on the Hugging Face repository page, which lists the Apache 2.0 license and the multimodal tags.

What an embedding is, in plain language

An embedding is a way of turning a piece of content, whether a sentence, a photo, a video clip or a snippet of audio, into a short list of numbers called a vector. Content with similar meaning ends up represented by similar lists of numbers. A search application can then compare the vector of your question with the vectors of your files and find the closest matches, without needing to understand either in human terms.

Think of it as placing every item on a very large map where proximity means similarity. A text query about a birthday party and a home video of a birthday party land near each other on that map, even though one is words and the other is pixels and sound. Traditional search relies on filenames or typed keywords; embedding search relies on meaning.

The novelty of EmbeddingGemma 2 is that it uses one shared map for all four media types. Earlier small embedding models typically produced one map for text only, requiring separate specialist models for images or audio. According to Google's announcement, this model natively maps text, code, images, video and audio into a single unified embedding space, so a spoken sentence and a video clip can be compared directly. Google describes a use case of locating a specific video clip from a voice memo, or searching hours of audio recordings with a text query.

Why on-device and open licensing matter

The privacy consequence comes from where the computation happens. When search runs on the device, the media being searched, the recordings, photos and videos, never leaves the phone. That matters for sensitive material: medical voice notes, family videos, confidential documents. It also matters for connectivity and cost, since on-device search works offline and produces no per-query cloud fees.

The accessibility consequence comes from the license and the size. Apache 2.0 is a commercially permissive license, meaning a two-person startup can embed the model in a shipping product without negotiating fees or usage restrictions. And because the model is small by modern standards, it does not require data-center hardware. Google states it can run on consumer devices such as phones and laptops.

A third consequence is modularity. Google reports that the model combines a 270 million parameter text backbone with separately loadable vision (170 million) and audio (300 million) encoders, so a developer building a text-only code search tool can skip the heavier encoders entirely and save memory.

One independent observation: on the Hugging Face repository, the listed parameter count of the released checkpoint is 744,371,512 in the BF16 safetensors metadata, which is consistent with Google's rounded "740 million" description rather than contradicting it. The repository page also shows quantized variants already published by third parties, indicating early ecosystem uptake, though download counts at retrieval were modest and the 20-million-download figure Google cites refers to the original EmbeddingGemma, the predecessor model.

The performance claims, attributed where they belong

All quality and memory figures below are Google's own reporting and should be read as vendor claims, not independently confirmed results.

According to the announcement and the model card material mirrored on the Hugging Face repository page, EmbeddingGemma 2 improves code performance on MTEB Code from 68.76 to 78.68, a 9.92-point gain over EmbeddingGemma 1, while matching its predecessor's multilingual text score on MTEB (61.36 versus 61.15). New benchmark results are listed for images (MIEB lite, 64.64), video (MMEB v2 Video, Hit@1 of 50.67), and audio (MSEB Retrieval MRR@10 of 69.54, MAEB 49.39). Google claims the model outperforms some specialist models more than twice its size.

Google also reports that the model supports Matryoshka Representation Learning, meaning output vectors can be truncated from 768 dimensions down to 512, 256 or 128, cutting vector storage by up to 6x with, per the repository documentation, minimal quality impact down to 256 dimensions. It features an 8K token context window, four times larger than the predecessor, which Google says allows processing up to 5.5 minutes of audio, 29 images or 58 video frames in one pass on local hardware.

Two cautions a reader should hold in mind. First, benchmark leadership is claimed "for its size" among sub-1B embedders, a specific and narrower comparison than leading among all models. Second, the RAM figures assume quantization on a specific device, the Pixel 11 Pro, so real-world results on other hardware may differ.

Verification status and the larger trend

Independent verification status for this article is as follows. The Google DeepMind blog post of October 6, 2026 was retrieved in full, including its dateline, author credits and headline claims. The Hugging Face weights repository google/embeddinggemma-2 was retrieved and confirms the release, the Apache 2.0 license tag, the multimodal capability tags, a last-modified timestamp of October 6, 2026, and checkpoint metadata consistent with the announced parameter count. The dedicated model card page at ai.google.dev could not be retrieved during preparation because the host was outside the research tooling's allowlist; the repository page's model card content, which covers architecture, benchmarks and usage guidance, was used instead. Readers should treat this as a minor sourcing gap rather than a contradiction.

On the broader picture, this release fits a pattern in which increasingly capable models are pushed down into openly licensed, small-footprint form factors. The consequence, if the vendor's claims hold up in practice, is that capabilities which previously implied sending personal media to a cloud service become feasible locally, offline, at no per-use cost, using weights anyone can download and inspect. That shift raises the baseline for what privacy-respecting personal software can do, and it lowers the entry barrier for developers outside large companies to build with genuinely multimodal AI.