EmbeddingGemma 2

Developers can use it to build on-device search across text, images, and audio on supported Macs.

Availability

Download from Hugging Face; versions 1.0 and 1.1 remain available as earlier revisions.

Pricing model

Apache 2.0 license

Deployment

macOS 27 or later on Apple silicon.

Privacy

Not established

Licence

Apache-2.0; Gemma Prohibited Use Policy pointer listed in NOTICE.

• Supports more than 100 languages. ^[1]
• It is designed to run on consumer hardware such as mobile devices and laptops. ^[1]

  Designed to run on consumer hardware such as mobile devices and laptops, EmbeddingGemma 2 delivers low-latency semantic representations for on-device applications, like search, retrieval-augmented generation (RAG), classification, and clustering.

• EmbeddingGemma 2 maps text, code, images, video, and audio into a shared 768-dimensional space. ^[2]

  Based on Gemma 4, this sub-1B model maps text, code, images, video, and audio into a unified 768-dimensional space.

• The model has configurations ranging from 270M parameters for text and code to 740M parameters for all modalities. ^[2]

  Its modular architecture lets you load only what you need, scaling from 270M parameters for text and code up to 740M parameters for all modalities.

• Matryoshka Representation Learning enables truncating embeddings from 768 dimensions to 128 dimensions. ^[2]

  Flexible vector storage: Matryoshka Representation Learning (MRL) enables dynamic truncation from 768 dimensions down to 128 dimensions, cutting vector database storage requirements while retaining much of the original quality.

• Model weights are available on Hugging Face. ^{[2] [3]}

  Download the Weights: Access the weights of the model (released under the Apache 2.0 license) directly on Hugging Face.

• The model can be run using sentence-transformers, Transformers, vLLM, SGLang, MLX, Ollama, LMStudio, and LiteRT. ^[2]

  Use Your Favorite Development Tools: Run the model efficiently using vLLM , Hugging Face Transformers , sentence-transformers , SGLang , MLX , Ollama , LMStudio , and LiteRT .

• EmbeddingGemma 2 scores 14% higher than EmbeddingGemma 1 on MTEB (Code), according to the announcement. ^[2]

  EmbeddingGemma 2 scores 14% higher than EmbeddingGemma 1 on MTEB (Code), adds image, video, and audio retrieval while retaining the accuracy on multilingual text of EmbeddingGemma 1

• Text, image, and audio embeddings share a 768-dimensional space. ^[4]

  1.1 the image encoder and since 1.2 the audio encoder, in one shared 768-d space.

• Audio embedding supports clips up to 30 seconds. ^[4]

  image-smoke-reference.json, preprocess-cases.json, check_preprocess.py, audio/ (16 generated clips

• Audio encoding can run on the Neural Engine or GPU. ^[4]

  - **Two audio routes**: the audio encoder runs on the Neural Engine (default) or on the GPU, set independently of

• The package can be downloaded from the EverPaper Hugging Face repository. ^[4]
• The package requires macOS 27 or later on Apple silicon. ^[4]

  embeddings) converted to Apple's **Core AI** runtime for **macOS 27 on Apple silicon**: the text encoder, since

• The release is licensed under Apache-2.0 and points to the Gemma Prohibited Use Policy. ^[4]

• Edit generated images through conversation with Gemini Nano Banana 2.1 Daily brief 2026-10-06