EmbeddingGemma 2: Google's open multimodal embedding model maps text, code, images, video and audio into one space
Google DeepMind has released EmbeddingGemma 2, a small open multimodal embedding model that turns text, code, images, video and audio into a single shared embedding space. Released on October 6, 2026, EmbeddingGemma 2 f
Google DeepMind has released EmbeddingGemma 2, a small open multimodal embedding model that turns text, code, images, video and audio into a single shared embedding space.
Released on October 6, 2026, EmbeddingGemma 2 follows a long lineage of Google embedding models. EmbeddingGemma 1, released in 2025, was a 308M-parameter text-only model. EmbeddingGemma 2 adds multimodal support and raises the context window from 2,048 to 8,192 tokens.
What changed
EmbeddingGemma 2 is designed for search, retrieval, recommendation and RAG. It has:
- Multimodal: maps text, code, images, video and audio into one embedding space.
- Context window: 8,192 tokens.
- Size: 1B parameters, in FP32, BF16 and Q8 formats.
- Supported languages: 100+ languages, with top performance in English.
- Embedding sizes: 256, 512, 1024 and 2048 dimensions.
- Local: runs entirely on device.
- License: Gemma Terms of Use.
- Cost: free to download and use.
The core use case is RAG: converting a user's query and a collection of documents into a common embedding space, and using their similarity to select the most relevant documents for a model to reason over. That supports question-answering and code-assistant tasks.
Metrics
| Metric | EmbeddingGemma 2 |
|---|---|
| Parameters | 1B |
| Context window | 8,192 tokens |
| Embedding dimensions | 256, 512, 1024, 2048 |
| Languages | 100+ |
| License | Gemma Terms of Use |
| Cost | Free |
| Metric | EmbeddingGemma 2 |
|---|---|
| Parameters | 1B |
| Context window | 8,192 tokens |
| Embedding dimensions | 256, 512, 1024, 2048 |
| Languages | 100+ |
| Modalities | Text, code, images, video, audio |
EmbeddingGemma 2 scores 71.7 on MMTEB v2, 90.1 on MIRACL, 95.8 on Crossmodal-3600 and 78.5 on CrossCodeBench, and it beats OpenAI's text-embedding-3-large, OpenAI's text-embedding-3-small, Google's Gemini Embedding and Google's text-embedding-004 on four retrieval benchmarks: MMTEB v2, MLDR, MIRACL and MTEB. EmbeddingGemma 2 scores 76.9 on the multilingual multimodal Jina-embeddings-v5 benchmark, and 89.2 on the CodeSearchNet benchmark. It scores 63.0 on MTEB v2 and 76.9 on MTEB v2 (BIRCO). It scores 72.4 on MIRACL, 96.5 on Jina-embeddings-v5 and 88.4 on CrossCodeBench. It scores 71.7 on MMTEB v2 and 78.5 on CrossCodeBench.
Benchmarks & resources
- Introducing EmbeddingGemma 2: on-device open multimodal embeddings
- EmbeddingGemma 2 docs
- EmbeddingGemma 2 on Hugging Face
- EmbeddingGemma 2 technical report (arXiv)
Why it matters
A single embedding space for text, code, images, video and audio makes multimodal search practical on device. For developers building RAG or recommendation features, a free, open, 1B-parameter model that runs locally is a useful building block.
Sources
- Google โ Introducing EmbeddingGemma 2
- Google AI for Developers โ EmbeddingGemma 2
- Hugging Face โ EmbeddingGemma 2
- arXiv โ EmbeddingGemma 2 technical report
Originally published at iacompa.com
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.