HomeAIAI ModelsGoogle's EmbeddingGemma 2 Puts Open Multimodal Embeddings on Phones

Google’s EmbeddingGemma 2 Puts Open Multimodal Embeddings on Phones

Google DeepMind has released EmbeddingGemma 2, an open embedding model built to put text, images, audio and video into one vector space while staying small enough to run on phones. Most embedding models still work with text alone.

On 6 October 2026, Google’s developer blog announced EmbeddingGemma 2: a 740-million-parameter embedding model under Apache 2.0, built on the Gemma 4 architecture, and described as natively mapping combinations of text, images, audio and video into a unified embedding space. [1]

What Google says the model is

According to the company blog (Research Engineers Sahil Dua and Henrique Schechter Vera, Google DeepMind), EmbeddingGemma 2 is:

  • Released under a commercially permissive Apache 2.0 licence.
  • Modular: as little as about 270M parameters for text-only workloads, with optional vision (170M) and audio (300M) encoders for full multimodal support.
  • Storage-efficient via Matryoshka Representation Learning (MRL), truncating output vectors from 768 dimensions down to 512, 256 or 128—claimed as up to 6Ă— storage reduction for local vector databases.
  • Optimised for on-device use: with quantisation, Google reports about 191MB active RAM for text-only weights and about 567MB for the full multimodal model on a Google Pixel 11 Pro.
  • Extended context: an 8K token context window (stated as 4Ă— EmbeddingGemma 1), said to allow up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations on local hardware. [1]

The same post says the model is built from the same technology as Gemini Embedding models and shares Gemma 4’s text tokenizer and audio encoder, which Google argues lowers combined memory when paired with Gemma 4 for on-device RAG. [1]

EmbeddingGemma 2 benchmark gains are Google’s own figures

Google reports a 9.92-point improvement on MTEB Code, from 68.76 to 78.68, while saying multilingual text performance matches EmbeddingGemma. It also claims leading scores among sub-1B multimodal embedders on benchmarks such as MTEB Code and MAEB (Massive Audio Embedding Benchmark), and that the model matches or outperforms many larger models across text, vision and audio tasks. [1]

Those figures are company-reported. TSN has not re-run MTEB, MAEB or on-device memory measurements.

How Google wants developers to use it

The blog points to on-device semantic search, video-moment finding, multimodal RAG with Gemma 4, and MediaPipe decision/routing tasks. Weights are said to be on Hugging Face and Kaggle, with Gemini Enterprise Agent Platform Model Garden availability “coming soon,” plus tooling mentions spanning LiteRT, MediaPipe, transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio and Qdrant. [1]

Google also cites more than 20 million downloads of the earlier EmbeddingGemma text model as community context—not as a quality proof for version 2. [1]

Update, 9 October 2026: Google’s Foresight meeting app for Macs

Google has put EmbeddingGemma 2 inside a Mac app. In a Google Developers Blog post dated 6 October, the Google AI Edge team announced Google AI Edge Foresight, which it calls “an experimental app that brings a context-aware meeting companion directly to your device. Powered by fully local AI processing” [2]. Google says it uses system audio and the microphone, so it “works out-of-the-box with any meeting platform, even completely offline” [2]. Wider coverage followed on 7 and 8 October [3][4].

What it does, in Google’s words. Foresight is “Powered by EmbeddingGemma 2 and Gemma 4 models”. Google lists three things: it “Enriches manual shorthand notes with details retrieved in real-time from your current meeting conversation”; it can “find different types of media (images, documents, transcripts, and notes) using natural language”; and it keeps “sensitive information” on the device, “with no cloud subscription costs” [2]. That cross-media search is the practical use of the shared vector space described above: EmbeddingGemma 2 turns notes, transcripts and files into vectors, and Gemma 4 writes the answers.

What reporters add. TechCrunch says users can chat with “the Gemma 4-powered assistant” and add PDFs, Google Docs, Microsoft Office formats, plain text, Markdown and web bookmarks to a knowledge base that the app draws on during meetings [3]. heise reports that Google says inference runs “100 percent locally on your device”, that the app is a direct download rather than an App Store app, and that it runs on Apple Silicon Macs only; “Intel Macs are not supported” [4]. TechCrunch frames it as a competitor to the AI note-taker Granola; that comparison is TechCrunch’s, not Google’s [3].

Other news in the same post. Google says EmbeddingGemma 2 is coming to Android’s ML Kit “in the coming weeks”, with NPU acceleration where available, and that MediaPipe Tasks is adding support for it [2].

Claims Google has not yet shown independently

  • “Best-in-class for its size” is a company framing; independent third-party leaderboards are not cited in the source used here beyond Google’s own summary.
  • Apache 2.0 covers the released weights as described by Google; downstream app compliance (privacy, on-device data handling) still sits with the developer.
  • Pixel 11 Pro RAM figures are Google measurements under quantisation; other devices and runtimes may differ.
  • Model Garden timing remains “coming soon.”
  • Whether Foresight becomes a finished product. Google calls it “experimental” [2].
  • Whether nothing leaves the device. “Fully local” processing is Google’s claim; TSN has not tested the app or its network traffic [2][4].

Update, 3 November 2026: EmbeddingGemma 2 reaches Android’s ML Kit as a preview

Google’s promised Android route has arrived in preview form. ML Kit’s release notes list a new genai-embedding library, version 1.0.0-alpha1, dated 26 October 2026 [5]. Google’s documentation says the Embeddings API uses EmbeddingGemma 2, managed by Android AICore, for text and image inputs [6]. It is a preview “not subject to any SLA or deprecation policy”, and the Google account on the device must be in the AICore Developer Preview programme [6]. The ML Kit version has a 256-token limit, with each image counting as 140 tokens, well short of the model’s 8K context [6]. Audio and video input are not listed for the ML Kit API.

EmbeddingGemma 2 on-device search trades size for scope

The interesting part of EmbeddingGemma 2 is less its benchmark scores than its shape. A model that can drop to about 270M parameters for text, add encoders only when needed, and shrink its vectors to 128 dimensions is designed around the limits of a phone’s memory and storage rather than a data centre. That matters for privacy: search over personal photos, notes or meeting audio can run without sending them to a server, the same on-device approach behind Google DeepMind and Singapore’s phone-based Gemma scam detector.

The ML Kit preview shows the cost of that convenience. Android developers get a managed model with no APK bloat, but on Google’s terms: text and images only, a 256-token window and an opt-in preview programme. Developers who want the full audio and video range, or the 8K context, still need to ship the weights themselves through LiteRT or another runtime. Open embedding models are multiplying, from Kakao’s Kanana-2 models for Korean search to Ai2’s satellite embeddings, so the Apache 2.0 licence and a single Gemma 4 tokenizer are Google’s real advantages, provided its memory and quality figures hold up in independent tests.

Sources

Sources

  1. Google (Sahil Dua & Henrique Schechter Vera, Google DeepMind), “EmbeddingGemma 2: an open, lightweight multimodal embedding model,” 6 October 2026 — https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/
  2. Google AI Edge Team and Android ML Team, “Bring multimodal semantic search to the edge with EmbeddingGemma 2”, Google Developers Blog, dated 6 October 2026 (company announcement; Foresight section). https://developers.googleblog.com/en/google-ai-edge-with-embeddinggemma-2/ (app page: https://developers.google.com/edge/foresight)
  3. Ivan Mehta, “Google releases a new local-first Granola competitor”, TechCrunch, 8 October 2026, 14:28 BST (trade press; Granola comparison is TechCrunch’s). https://techcrunch.com/2026/10/08/google-releases-a-new-local-first-granola-competitor/
  4. Ben Schwan, “Google’s macOS app: Local AI to capture notes and work with files”, heise online (English edition of a Mac & i article), 7 October 2026, 22:02 BST (trade press, quoting Google). https://www.heise.de/en/news/Google-s-macOS-app-Local-AI-to-capture-notes-and-work-with-files-11479591.html
  5. Google, “ML Kit release notes”, Google for Developers, genai-embedding 1.0.0-alpha1 listed as last updated 26 October 2026, read 3 November 2026. https://developers.google.com/ml-kit/release-notes
  6. Google, “Semantic search using the Embeddings API”, ML Kit documentation for Android, read 3 November 2026. https://developers.google.com/ml-kit/genai/embedding/android

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

  1. WashU Makes Edible Fungal Protein From Waste CO2
    AI Science & Health

More in this category

Latest stories

  1. WashU Makes Edible Fungal Protein From Waste CO2
    AI Science & Health

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.