What changed: Google DeepMind introduced EmbeddingGemma 2, which natively maps text, code, images, video and audio into a unified embedding space. Built on the Gemma 4 architecture and released under the Apache 2.0 license, the model has 740 million parameters, improved MTEB Code score from 68.76 to 78.68 (a 9.92-point gain), and extended its context window to 8K tokens, 4x larger than EmbeddingGemma 1.
Who is affected
The source names developers and integration partners (Hugging Face, Kaggle, Qdrant, Ollama, LMStudio among others); no specific country or sector restriction is stated.
What it means for your business
For businesses building on-device search or RAG pipelines, this means a new open-source model, released under the Apache 2.0 license, that can now process image, audio and video alongside text in a single download.
What to watch
No action is required right now. Businesses using AI for product search, support, or content matching can have their technical teams evaluate the model's fit for their own infrastructure.
Timing
The model was announced on October 6, 2026 and is downloadable via Hugging Face and Kaggle; the source gives no separate effective date.
Source: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/