Google releases EmbeddingGemma 2, an open model that maps text, images, audio and video into one search space

What launched

Google DeepMind released EmbeddingGemma 2 on 6 October 2026.¹ It is an open embedding model: it turns text, code, images, audio and video into lists of numbers, called vectors, in one shared space, so an app can search across all of them at once.¹²

Key specs

• Size: 740 million parameters for the full model. Text and code alone need 270 million, and developers can load only the parts they need.¹²
• Built on: the Gemma 4 architecture.¹
• Output: 768-dimensional vectors by default, which can be shortened to 512, 256 or 128 for up to six times less storage.¹²
• Input length: 8,192 tokens, four times the first EmbeddingGemma. Google says that covers about 5.5 minutes of audio, 29 images or 58 video frames.¹
• Licence: Apache 2.0, which allows commercial use.¹²

On a phone

Google says that, when quantised, the text-only model uses about 191MB of active memory on a Pixel 11 Pro, and the full multimodal model about 567MB.¹

Benchmark claims

Google says the model’s MTEB Code score rose from 68.76 to 78.68 compared with the first version, and that it leads multimodal embedding models under 1 billion parameters on code and audio benchmarks.¹ The developer guide describes the code gain as 14%.² These are Google’s own results.

Where to get it

Weights are on Hugging Face (google/embeddinggemma-2) and Kaggle.¹ Google lists support in tools such as sentence-transformers, Transformers, vLLM, Ollama and LM Studio.¹² It says the first EmbeddingGemma passed 20 million downloads.¹

Who it is for

Developers building search, recommendations or retrieval-augmented generation (RAG) that must run on a device or on modest hardware, especially apps that mix text with images, audio or video.