EmbeddingGemma 2 is Google’s new multimodal on‑device embedding model built on the Gemma 4 architecture and released under an Apache 2.0 license. The model contains 740 million parameters and is designed to compute embeddings locally across text, code, images, video and audio within a single shared embedding space.
Why it matters
The original EmbeddingGemma exceeded expectations with more than 20 million downloads and has been used in on‑device search tools and privacy‑first retrieval‑augmented generation (RAG) pipelines. EmbeddingGemma 2 extends that capability by unifying multiple modalities, enabling scenarios such as locating a specific video clip from a voice memo or searching hours of audio with a text query — all processed by one native multimodal model.
Key technical highlights
- Parameters: 740 million.
- License: Apache 2.0.
- Architecture: built on Gemma 4; shares Gemma 4’s text tokenizer and audio encoder.
- Extended context: an 8K token context window (about 4× larger than EmbeddingGemma 1), enabling local processing of up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations of these.
Performance and efficiency
- Size‑to‑performance: on benchmarks such as MTEB (text and code) and MAEB (audio), EmbeddingGemma 2 achieves leading scores among sub‑1B multimodal embedders and matches or outperforms many larger models on text, vision and audio tasks.
- Code performance: on the MTEB Code benchmark, EmbeddingGemma 2 improves by 9.92 points over its predecessor (from 68.76 to 78.68), making it suitable for local codebase indexing, semantic code search and retrieval for coding agents.
Modular design and storage optimizations
- Modular configurations: the model can run with as little as 270M parameters for text‑only workloads; optional encoders add ~170M for vision and ~300M for audio to enable full multimodal support.
- Matryoshka Representation Learning (MRL): output vectors can be truncated dynamically from 768 dimensions down to 512, 256 or 128, which can yield up to a 6× storage reduction for local vector databases and memory usage.
On‑device performance and memory footprint
- Quantized measurements on a Google Pixel 11 Pro show active RAM needs of approximately ~191 MB for text‑only weights and ~567 MB for the full multimodal model. These figures reflect optimizations for constrained hardware.
Use cases and tooling
EmbeddingGemma 2 enables robust on‑device semantic search, routing and retrieval, reducing data transfer, lowering latency and improving privacy. When paired with generative models such as Gemma 4, it supports on‑device RAG pipelines that handle complex multimodal data while sharing tokenizer and audio encoder components to reduce combined memory footprint.
Examples and integrations in practice include searching a media library by text or image for semantically similar items, locating moments in video with text or audio queries, and combining EmbeddingGemma 2 for local file retrieval with Gemma 4 for contextual reasoning.
Tooling and deployment options: model weights are available on Hugging Face and Kaggle; optimized, on‑device variants can be found via LiteRT Community resources. Developers can integrate the model with Google AI Edge, MediaPipe and LiteRT for native on‑device deployment, or run models in the browser with transformers.js or WebGPU. Supported serving and development tools include transformers, sentence‑transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio; embedding vectors can be stored in vector databases such as Qdrant.
Fine‑tuning and resources
Community guidance for fine‑tuning (for example, materials published by Unsloth) and a detailed EmbeddingGemma 2 model card with evaluation metrics are available for developers looking to adapt the model to their use cases.
Availability
Model weights and on‑device optimized versions are accessible on Hugging Face and Kaggle, with availability in the Gemini Enterprise Agent Platform Model Garden expected soon. Further developer guides, documentation, and examples for inference and fine‑tuning are published through Google AI Edge channels and accompanying documentation.
