Google DeepMind releases EmbeddingGemma 2, an open multimodal embedding model

Google DeepMind released EmbeddingGemma 2, an embedding model, on October 6, 2026. An embedding model turns a piece of content into a list of numbers, a vector, that represents its meaning; content with similar meaning gets similar vectors. This lets search, document retrieval and grouping of similar items work without relying on keyword matches. The new model maps text (including code), images, video and audio into one shared space, so a written query can find a moment in a video or an audio recording. The weights are openly available under the Apache 2.0 license.

According to Google's announcement, the model is designed to run on devices such as phones and laptops without sending data elsewhere. Example uses include searching a photo and video library on the device, local RAG (retrieval-augmented generation) pipelines that answer from your own files, semantic code search and content classification. EmbeddingGemma 2 follows last year's text-only EmbeddingGemma and is built on the Gemma 4 architecture.

A modular design that loads only what you need

According to the model card, EmbeddingGemma 2 has 740 million parameters in total: 270 million for the text part, 170 million for the vision encoder and 300 million for the audio encoder. Because the vision and audio encoders load separately, a text-only application can use just the 270-million-parameter part; text plus images needs about 440 million parameters, and text plus audio about 570 million. Google says a quantized version uses about 191 MB of memory for text only and about 567 MB for the full multimodal model on a Pixel 11 Pro phone.

The model understands more than 100 languages and produces 768-dimensional vectors. Thanks to a training method called Matryoshka representation learning, these vectors can be shortened to 512, 256 or 128 dimensions; shorter vectors take less storage and speed up search. The context window, the amount of input the model reads at once, is 8K tokens. According to the model card, that fits about 29 images, 58 video frames or up to 5.5 minutes of audio on their own. Text, images and video can be mixed in a single input and turned into one vector, so a product listing can be indexed together with its description, photos and demo video.

Benchmarks shared by Google

According to Google's published results, the model raises the original EmbeddingGemma's score on the MTEB Code retrieval benchmark from 68.76 to 78.68, while its multilingual MTEB text score moves from 61.15 to 61.36. The company says it posts the top scores on MTEB Code and the MAEB audio benchmark among multimodal embedding models under 1 billion parameters, and that it beats some specialist models more than twice its size on image, video, document and audio tasks.

Options for running it locally

The weights are published on Hugging Face as google/embeddinggemma-2 and on Kaggle. In Python, it takes a few lines with the sentence-transformers library: after pip install -U sentence-transformers transformers, the model loads with SentenceTransformer("google/embeddinggemma-2") and uses separate task prefixes for search queries and documents. According to the announcement, Ollama, llama.cpp (GGUF files), LM Studio, MLX, vLLM and SGLang are supported as well. The model also runs in the browser with transformers.js and WebGPU, and on Android devices through Google AI Edge's MediaPipe and LiteRT. The announcement also points to an Unsloth fine-tuning guide and use with the Qdrant vector database. The model card recommends bfloat16 or float32 for inference; in float16 the model can return broken results without raising an error.

The model card says the model supports more than 100 languages but may not perform equally well in all of them, so it is worth testing on your own examples before using it with non-English content. Shortening vectors to 128 dimensions noticeably lowers quality on multimodal tasks. The model does not go through the safety tuning applied to generative models, so safeguards such as filtering harmful content are left to the developers building on it. Availability in Google's Gemini Enterprise Agent Platform Model Garden has been announced as coming soon. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Google: EmbeddingGemma 2 announcement, Hugging Face: google/embeddinggemma-2 model card.