Native multimodal embeddings
Maps text, code, images, video, and audio into one shared embedding space, supporting semantic comparisons across different media types.
EmbeddingGemma 2
EmbeddingGemma 2 is an open, lightweight multimodal embedding model from Google DeepMind for private local search, semantic retrieval, code indexing, and on-device RAG.
Overview
EmbeddingGemma 2 is an open, lightweight multimodal embedding model from Google DeepMind for private local search, semantic retrieval, code indexing, and on-device RAG.
It converts combinations of text, code, images, video, and audio into a unified embedding space, allowing applications to find semantically related content across media types rather than relying only on exact keywords. The model is designed for privacy-oriented, on-device workflows, with a modular architecture, configurable vector dimensions, and an 8K-token context window for constrained local environments.
Key features
On-device inference
Uses a 740-million-parameter architecture designed for local inference, with quantized memory requirements reported as approximately 191 MB for text-only weights and 567 MB for the full multimodal model on a Google Pixel 11 Pro.
Modular model design
Supports a text-only configuration of about 270 million parameters, with optional 170-million-parameter vision and 300-million-parameter audio encoders for multimodal workloads.
Extended context
Provides an 8K-token context window, supporting up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations directly on local hardware.
Flexible vector sizes
Matryoshka Representation Learning allows vectors to be truncated from 768 to 512, 256, or 128 dimensions, reducing local vector-database and memory requirements by up to six times according to the source.
Open developer ecosystem
The model is released under the Apache 2.0 license and can be used with Google AI Edge MediaPipe, LiteRT, browser tooling such as transformers.js and WebGPU, and several model-serving frameworks listed by Google.
Use cases
Private media search
Embed files, images, recordings, and other personal media locally, then retrieve related items with a text or image query without sending the source content to a remote service.
Audio and video moment retrieval
Search long recordings or video collections by entering a text description or using an audio query to locate semantically related moments.
Codebase indexing
Index local repositories and support semantic code search or coding-agent retrieval. The source reports a 9.92-point improvement over the first EmbeddingGemma on MTEB Code.
On-device RAG
Use EmbeddingGemma 2 for local retrieval and pair it with Gemma 4 for contextual reasoning in a multimodal, on-device RAG pipeline.
Local classification and routing
Use multimodal context in MediaPipe Decision Task API workflows for classification, routing, and predictive decision engines.
Pros and Cons
Pros
- Combines text, code, image, video, and audio embeddings in one shared space.
- Designed for offline or local processing, which can reduce data transfer and support privacy-sensitive workflows.
- Offers modular encoders and selectable vector dimensions for different resource budgets.
- Supports local search, multimodal retrieval, and on-device RAG workflows.
- Released under the commercially permissive Apache 2.0 license.
Cons
- The source provides launch-level information but does not specify detailed API behavior, supported operating-system versions, latency benchmarks, or operational limits.
- Full multimodal inference requires more resources than the text-only configuration, with the source reporting substantially higher active RAM for the full model.
- Pricing and service terms are not provided; the model’s Apache 2.0 release does not by itself describe every deployment or hosting consideration.
FAQ
What is EmbeddingGemma 2?
EmbeddingGemma 2 maps text, code, images, video, and audio into a shared embedding space so applications can compare and retrieve different media types using semantic similarity.
Can EmbeddingGemma 2 run on-device?
The model can run on local hardware and is designed for on-device inference. Google reports approximately 191 MB of active RAM for quantized text-only weights and approximately 567 MB for the full multimodal model on a Google Pixel 11 Pro.
How can developers get started with EmbeddingGemma 2?
Developers can download model weights from Hugging Face and Kaggle. The source also lists Google AI Edge MediaPipe, LiteRT, transformers.js, WebGPU, transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio as supported development or serving tools and environments.
What embedding dimensions does EmbeddingGemma 2 support?
The model supports output vectors that can be dynamically reduced from 768 dimensions to 512, 256, or 128 dimensions using Matryoshka Representation Learning. The source describes this as enabling up to a sixfold reduction in local vector-database storage and memory usage.
How much does EmbeddingGemma 2 cost?
The launch article does not provide pricing information. It states that the model is released under the commercially permissive Apache 2.0 license, but users should consult the model card and accompanying documentation for applicable terms and usage details.
Quick Facts
- Category
- Developer tool; AI embedding model
- Developer
- Google DeepMind
- Parameters
- 740 million full model
- Context window
- 8K tokens
- License
- Apache 2.0
- Model sources
- Hugging Face and Kaggle
EmbeddingGemma 2 Alternatives
ByteAsk
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repos and verifies changes with compilers, debuggers, sanitizers, and tests.
Epherem
Epherem scans web apps built with Lovable, Bolt, v0, Replit, Cursor, and Claude Code for security flaws in source code and dependencies, without running the app.
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads in Firecracker micro-VMs with private networking and SDK, CLI, or MCP control.
GitBot
GitBot is a local browser workspace for reusable coding bots powered by Claude Code, Codex, or OpenCode, with persistent threads and configurable permissions.
hob
hob is an independent workspace for coding agents, with local control over sessions, terminals, history, routing, and follow-up work.
Ably Chat
Ably Chat is a chat API platform for custom realtime chat apps, with rooms, typing indicators, presence, reactions, message updates and usage-based pricing.