Artificial Intelligence
Google launches EmbeddingGemma 2 multimodal embedding model
Google launched EmbeddingGemma 2, a 740 million-parameter multimodal embedding model, and said it could run on-device across code, images, video and audio under an Apache 2.0 licence.
- Google said the new model extended EmbeddingGemma from text embeddings to code, images, video and audio in one shared embedding space.
- The company said EmbeddingGemma 2 had 740 million parameters and was released under Apache 2.0.
- Google said the earlier EmbeddingGemma had passed 20 million downloads.
- The company said the model achieved leading results among sub-1 billion multimodal embedders on benchmarks including MTEB Code and MAEB, and that it improved code scores from 68.76 to 78.68.
- Google listed partners and tools including Hugging Face, Kaggle, Google AI Edge MediaPipe, LiteRT, transformers.js, WebGPU, Qdrant and Unsloth.
Google launched EmbeddingGemma 2 as an open multimodal embedding model and said it broadened the earlier EmbeddingGemma from text to code, images, video and audio. It said the first EmbeddingGemma had more than 20 million downloads.
The company said EmbeddingGemma 2 used the Gemma 4 architecture, carried 740 million parameters and was released under Apache 2.0. Google said the model was intended for on-device inference and could support tasks such as finding a video clip from a voice memo or searching audio by text.
Google said the model shared technology with Gemini Embedding models and could be combined with Gemma 4 in on-device retrieval-augmented generation pipelines. It also said the model used Matryoshka Representation Learning to let developers shorten output vectors to 768, 512, 256 or 128 dimensions.
In its announcement, Google said the model reached leading scores among sub-1 billion multimodal embedders on code, vision and audio tasks. It also said it reduced memory use on a Google Pixel 11 Pro to about 191MB for text-only weights and about 567MB for the full multimodal model.
Google said EmbeddingGemma 2 was available through Hugging Face and Kaggle, with later availability planned for Gemini Enterprise Agent Platform Model Garden. It also named Google AI Edge MediaPipe, LiteRT, transformers.js, WebGPU, Qdrant and Unsloth among the tools and partners it expected developers to use.
Named in this story
Companies
- announced the model and made the performance claims
- Qdrant
- was named as storage for embedding vectors
- Unsloth
- was cited for fine-tuning guidance
Organisations
- Hugging Face
- was named as a place to download the model weights
- Kaggle
- was named as a place to download the model weights
- LiteRT Community
- was named as a source for on-device models
Products and systems
- EmbeddingGemma 2
- the multimodal embedding model Google launched
- EmbeddingGemma
- the earlier text embedding model Google said it released last year
- Gemma 4
- the architecture and paired model family used with EmbeddingGemma 2
- Gemini Embedding models
- the models Google said EmbeddingGemma 2 used the same technology as
- Gemini Enterprise Agent Platform Model Garden
- was said to be getting model availability later
- Google AI Edge MediaPipe
- was recommended for cross-platform on-device deployment
- LiteRT
- was recommended for custom model integration
- transformers.js
- was listed as a browser deployment option
- WebGPU
- was listed as a browser deployment option
- Google Pixel 11 Pro
- was used in Google’s memory-use example
How the source tells it
The piece is a promotional product announcement, written in a boosterish register with repeated superlatives and performance claims attributed to Google.
- vendor boosterism superlatives and comparative claims are used to present the model as leading its class
- hype marketing language describes the product as optimal, top-tier, and a new standard