EmbeddingGemma 2: Google DeepMind's new multimodal model for mobile devices

On October 6, 2026, Google DeepMind announced EmbeddingGemma 2 on its official blog, in a post signed by researchers Sahil Dua and Henrique Schechter Vera. The model has 740 million parameters in total and projects text, code, images, video and audio into the same vector space. Its modular architecture lets you switch on only the parts you need: the text-and-code section uses about 270M parameters, while the vision and audio encoders add 170M and 300M respectively. Every configuration shares the same vector space, which makes cross-modal semantic search straightforward.

The Apache 2.0 license, which Google calls “commercially permissive”, allows commercial use too. EmbeddingGemma 2 is built on the Gemma 4 architecture and supports an 8,192-token context, four times larger than version 1. That works out to roughly 5.5 minutes of audio, 29 images or 58 video frames per input. Native vectors have 768 dimensions, but thanks to Matryoshka Representation Learning they can be cut down to 512, 256 or 128 dimensions, shrinking storage by up to 6x. The Hugging Face model card flags a substantial drop in multimodal quality at 128 dimensions and recommends that size only for text-only workloads.

On performance, Google reports an MTEB Code score of 78.68, 9.92 points above the previous version, while the multilingual v2 benchmark is almost unchanged (61.36 versus 61.15). Other benchmarks at 768 dimensions include MIEB lite 64.64, MMEB v2 images 57.28, MMEB v2 visual documents 67.84, MMEB v2 video 50.67, MSEB retrieval 69.54 and MAEB 49.39. Google states that EmbeddingGemma 2 ranks first among sub-billion-parameter multimodal embedders on MTEB Code and MAEB. All of these scores, like the leaderboard claim, come from Google and its Hugging Face card and have not yet been independently verified.

According to Google, on a Pixel 11 Pro and in quantized form, the text part alone takes about 191 MB of active RAM and the full model about 567 MB. The figure is the company's own and covers a single device. The weights are available on Hugging Face and Kaggle, and the model is supported by many frameworks (transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio, transformers.js/WebGPU) as well as MediaPipe/LiteRT for on-device execution. The card warns that the model is pre-trained without post-training alignment, safety tuning or output moderation, so developers must put the necessary safeguards in place. It also advises against float16, which can produce NaNs or degraded embeddings without raising any error.

The card notes that performance may vary across the more than 100 supported languages: results specifically for Italian are not documented. Availability in the Model Garden of the Gemini Enterprise Agent Platform is announced as “coming soon”, still with no date.

— Pixie