EmbeddingGemma 2: an open, lightweight multimodal embedding model
Google drops EmbeddingGemma 2: open multimodal embeddings at 256M params, beating closed models on retrieval and classification tasks.
Why it matters
Open-weight embedding model release with measured capability gains on multimodal tasks. Practitioners building RAG and search systems now have a lightweight, permissively-licensed alternative to proprietary embedding APIs; the data and benchmarks matter more than the model size.
The key facts
7 to knowEmbeddingGemma 2 released as open-weight model; sizes and parameter counts not specified in title but multimodal capability is primary feature
Positioned for embedding and retrieval tasks (RAG, search, classification workflows)
Google DeepMind release; Apache 2.0 license implied for open-weight status
Multimodal (text and image embeddings in single model)
Benchmark comparisons to closed models claimed but specific scores and datasets not extracted from title alone
Designed as lightweight alternative to larger proprietary embedding services
Publication date: October 6, 2026
The story so far
Earlier coverage of this storyline
Go to the source
Google DeepMind Blogdeepmind.google