Google released EmbeddingGemma 2 on October 6, 2026, a small open embedding model that now goes well beyond text: it maps text, code, images, audio and video into one shared vector space. Built on the Gemma 4 architecture, the model has 740 million parameters and ships under the Apache 2.0 license, which allows commercial use. It is designed to run directly on phones and laptops, so the data never has to leave the device.
An embedding model turns content into numbers so software can search by meaning instead of exact keywords. It is the engine behind semantic search and RAG (retrieval-augmented generation), where an app fetches the most relevant documents before a language model answers. Google says the first, text-only EmbeddingGemma, released last year, passed 20 million downloads.
What’s new in EmbeddingGemma 2
- Natively multimodal: one search across media types. For example, finding a specific video clip from a voice memo, or searching hours of audio recordings with a typed question.
- Modular: text-only workloads need as little as 270 million parameters. The vision encoder (170 million) and audio encoder (300 million) are optional add-ons.
- 8K-token context: four times the first version. Google says that covers about 5.5 minutes of audio, 29 images or 58 video frames in a single input.
- Smaller indexes: with Matryoshka Representation Learning, vectors can be truncated from 768 dimensions to 512, 256 or 128, cutting local vector database storage by up to 6x.
- Low memory: quantized on a Pixel 11 Pro, Google says it needs about 191 MB of active RAM for text only and about 567 MB for the full multimodal model.
Performance, according to Google
Google calls EmbeddingGemma 2 the leading multimodal embedder under one billion parameters on benchmarks such as MTEB Code and MAEB (audio). The standout gain is in code: its MTEB Code score rose from 68.76 to 78.68, an improvement of almost 10 points, while multilingual text quality holds at the level of the first version. Google also says it beats some specialist models more than twice its size. These are Google’s own numbers, and independent evaluations haven’t been published yet.
Where to get EmbeddingGemma 2
The weights are available now on Hugging Face and Kaggle, with availability in the Gemini Enterprise Agent Platform Model Garden coming later. The model works with popular tools including transformers, sentence-transformers, Ollama, LM Studio, llama.cpp, vLLM and MLX, and it can run in the browser through transformers.js and WebGPU. For mobile apps there is MediaPipe and LiteRT support, and Unsloth has published fine-tuning guidance. Because it shares its tokenizer and audio encoder with Gemma 4, developers can run both together in one local RAG pipeline with a smaller combined memory footprint.
Why it matters
Most headline AI features, such as those in Gemini 4 Argon, run in the cloud. EmbeddingGemma 2 pushes the other way: search and retrieval over personal photos, recordings and files that stay on the device. For developers in the US and Europe, that means a free, open option for semantic search without per-call API fees, and a simpler privacy story under rules like the GDPR, since sensitive data doesn’t have to be sent to a third-party server. Follow AI News for more.
Sources
- Google, official blog (The Keyword): EmbeddingGemma 2: an open, lightweight multimodal embedding model, October 6, 2026, blog.google
