Google DeepMind Releases EmbeddingGemma 2, an Open Multimodal Embedding Model

Google DeepMind released EmbeddingGemma 2 on October 6, 2026, an open embedding model that maps text, code, images, video and audio into one shared vector space. The model has 740 million parameters and ships under the Apache 2.0 license, small enough to run on a phone, the company said.

An embedding model turns content into a list of numbers, a vector, so that software can find similar items by comparing those numbers. The first EmbeddingGemma, launched in September 2025, handled text only and passed 20 million downloads, according to Google. The new version extends the same idea to images, video and audio for search tools that run offline, and like its predecessor it ships as one of Google's open weight models.

What is EmbeddingGemma 2?

EmbeddingGemma 2 is a 740 million parameter model built on the Gemma 4 architecture that produces one 768 dimensional vector for text, code, images, video, audio or a mix of them, Google said in its announcement. Google says it is built from the same technology as its Gemini Embedding models.

The practical result is cross modal search. Google's examples include finding a video clip from a voice memo and searching hours of audio recordings with a typed query, all handled by one model instead of separate systems for each type of media.

The model card lists a context window of 8,192 tokens, four times the window of the first version, and support for more than 100 languages plus code. Each image costs 280 tokens, each video frame 140 tokens and each second of audio 25 tokens, according to the model card, which works out to about 29 images, 58 video frames or 5.5 minutes of audio in one 8,192 token input, the same limits Google gives in its announcement.

How small can it run?

EmbeddingGemma 2 is modular, so a text only setup needs about 270 million parameters, Google says. The vision encoder adds 170 million parameters and the audio encoder 300 million, and developers can switch either one off when they do not need it.

With quantization, a technique that stores the model's numbers at lower precision to save memory, Google reports that the model uses about 191 MB of active RAM for text only and about 567 MB for the full multimodal version on a Google Pixel 11 Pro. The Decoder reports that a query takes about 20 to 70 milliseconds in a browser through WebGPU.

The model also supports Matryoshka Representation Learning, a training method that lets developers cut vectors from 768 dimensions down to 512, 256 or 128. Google says this reduces storage for local vector databases by up to six times, and its developer guide says 256 dimensional vectors keep about 95% of full quality on image, video and speech retrieval.

How does it compare with the first EmbeddingGemma?

Google reports that EmbeddingGemma 2 scores 78.68 on the MTEB Code benchmark, up from 68.76 for the first version, while its multilingual text score stays roughly level. Google's own model card is the only source for these numbers so far.

Google also claims EmbeddingGemma 2 leads multimodal embedding models below 1 billion parameters on MTEB Code and on MAEB, an audio embedding benchmark, and that it beats some specialist models more than twice its size. The rivals appear in the charts of its announcement: Qwen3-Embedding-0.6B and Qwen3-Embedding-8B on code, jina-embeddings-v5-omni-small and LCO-Embedding-Omni-3B on images, and jina-embeddings-v5-omni-nano, BidirLM-Omni-2.5B and LCO-Embedding-Omni-7B on audio. On Google's own audio chart, those three models sit above EmbeddingGemma 2.

The model cards of those rivals on Hugging Face show where EmbeddingGemma 2 differs. Qwen3-Embedding-0.6B is a similar size under Apache 2.0 but handles text only. jina-embeddings-v5-omni-nano covers text, images, video and audio at about 1 billion parameters, but under a non-commercial CC BY-NC 4.0 license. LCO-Embedding-Omni-3B is Apache 2.0 and multimodal, with about 4.7 billion parameters listed. Of these four models, only EmbeddingGemma 2 combines text, images, video and audio with a commercial license and a size under 1 billion parameters.

Where can developers get it?

The weights are available on Hugging Face and Kaggle, with on device versions in the LiteRT Community on Hugging Face, according to Google. Google lists availability in its Gemini Enterprise Agent Platform Model Garden as coming soon, without a date.

Google says it worked with partners so the model runs in common tools at launch, including sentence-transformers, transformers.js, vLLM, SGLang, llama.cpp, Ollama, LM Studio and MLX, as well as the Qdrant vector database. Unsloth has published fine tuning guidance. Because EmbeddingGemma 2 shares its text tokenizer and audio encoder with Gemma 4, Google says the two can run together in one pipeline with a lower combined memory footprint.

The model card also lists limits: performance may vary across languages, and the model has no post training safety alignment, so Google leaves application level safeguards to developers.

What we don't know yet

• How EmbeddingGemma 2 performs in independent evaluations against other open multimodal embedding models.
• When the model will reach the Gemini Enterprise Agent Platform Model Garden.
• How quality holds up across the more than 100 supported languages, which the model card says may vary.