On Tuesday, October 6, 2026, Google DeepMind unveiled EmbeddingGemma 2, the second generation of its open, lightweight embedding model. The announcement, published on Google's blog by research engineers Sahil Dua and Henrique Schechter Vera, describes the new model as the most capable option for generating multimodal embeddings directly on the device, without relying on the cloud. Google released a technical guide for developers the same day.
An embedding is a numerical representation of a piece of content, a vector, that makes it possible to measure how similar two items are in meaning. It is the component that powers semantic search, recommendation systems and so-called retrieval-augmented generation (RAG), in which a language model looks up relevant documents before answering. The first version of EmbeddingGemma, launched last year, worked only with text and passed 20 million downloads.
The big change is native multimodality. EmbeddingGemma 2 represents text, code, images, video and audio, as well as interleaved combinations of all of them, in a single 768-dimension vector space. In practice, according to Google, that makes it possible to find a specific video clip from a voice memo or to search hours of audio recordings with a simple text query, all with a single model.
The model is built on the Gemma 4 architecture, is released under the Apache 2.0 license, which allows commercial use, and has 740 million parameters in total. Its design is modular: the text and code component needs only 270 million, split into a 130 million backbone and a 140 million embedder, while the vision encoder adds 170 million and the audio encoder 300 million. It can therefore be used in configurations of 270, 440, 570 or 740 million parameters.
That modularity brings an unusual advantage. Because every configuration shares the same vector space, a query made with the text-only setup can be compared against documents indexed with the full model. In addition, according to the developer guide, adding the image or audio encoders later does not require recomputing existing embeddings, which saves time and cost on large collections.
Context grows too. EmbeddingGemma 2 accepts 8,192 tokens per input, four times more than its predecessor. According to Google, that is equivalent to about 5.5 minutes of audio, 29 images or 58 video frames, or interleaved combinations of that content. By default, video is sampled at one frame per second and audio is processed as 16 kHz mono.
On performance, Google says the model keeps the strong multilingual text quality of the first version, with support for more than 100 languages, and improves markedly on code: on the MTEB Code benchmark it rises from 68.76 to 78.68 points, a gain of 9.92. The company claims it is best in class among multimodal models under one billion parameters on MTEB Code and MAEB, a comparison that comes from the vendor itself.
Storage savings are another central argument. Thanks to Matryoshka Representation Learning, vectors can be truncated from 768 to 512, 256 or 128 dimensions, for up to a sixfold reduction in storage. The guide gives an example: one million 768-dimension vectors in bfloat16 take up about 1.5 GB, compared with about 250 MB at 128 dimensions. At 256 dimensions, most of the quality on text and code is preserved.
That truncation is not free. According to the guide, at 256 dimensions the model keeps about 95% of its performance on image, video and speech retrieval, while at 128 dimensions it retains around 90% on text and code but about 75% on media content. Google recommends validating each configuration with your own data before moving it into production. It also warns that the model is not compatible with float16 and must run in bfloat16 or float32.
The on-device focus is explicit. With quantization, on a Google Pixel 11 Pro, EmbeddingGemma 2 needs about 191 MB of active memory for the text-only weights and about 567 MB for the full multimodal model. Because it shares the text tokenizer and audio encoder with Gemma 4, both models can run together in a single pipeline with a smaller total memory footprint, making it easier to build fully local RAG systems that send no data to external servers.
Google backed the launch with demos in its Google AI Edge Gallery app, such as instant search across a media library and a finder for moments inside videos, and with the Google AI Edge Foresight app, which combines local file retrieval with Gemma 4's reasoning. In the developer guide, Google indexed the codebase of Hugging Face's transformers library with the text-only setup and queried it through an agent.
Availability is broad from day one. The weights are on Hugging Face and Kaggle, and arrival in Model Garden is announced as coming soon. The model works with transformers, sentence-transformers from version 6.1.0, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio, as well as LiteRT and MediaPipe for mobile and transformers.js or WebGPU for the browser. Qdrant supports it as a vector database and Unsloth offers fine-tuning guides.
For developers, the appeal lies in the combination of an open license, a small footprint and a single space for every format. Companies handling sensitive documents can build internal search tools without sending content to the cloud, and builders of coding agents get a lightweight option for indexing repositories. The improvement on code points precisely at that use: semantic code search and retrieval for agents.
In conclusion, EmbeddingGemma 2 brings multimodal search down to a size that fits on a phone and releases it under a permissive license. The performance figures come from Google and will need to be checked against independent tests, and dimension truncation requires validating quality case by case. Even so, the launch reinforces Google's bet on more private, efficient AI that runs outside the data center.
