Google released EmbeddingGemma 2 on October 6, 2026, an open embedding model built on the Gemma 4 architecture that maps text, images, audio, video and code into a single shared embedding space. It has 740 million parameters in total; text-only use can run on as few as 270 million, with an optional 170 million-parameter vision encoder and a 300 million-parameter audio encoder. It is released under Apache 2.0 with weights on Hugging Face and Kaggle.
Google reports the MTEB Code score rose from 68.76 to 78.68, a 9.92-point gain over the first EmbeddingGemma, and calls it best in class among sub-1B multimodal embedders. The context window is 8K tokens, four times the original, and it can take up to 5.5 minutes of audio, 29 images or 58 video frames. Matryoshka representation learning lets its 768-dimension vectors be cut to 512, 256 or 128 dimensions for up to 6x less storage.
It is built to run on device: quantized on a Pixel 11 Pro it needs about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model, and it ships with support for MediaPipe, LiteRT, transformers.js, llama.cpp, Ollama, vLLM and other runtimes.
Why it matters: retrieval and search underpin most practical AI applications, and a small, permissively licensed model that embeds every common modality on a phone lets that layer run locally and privately instead of through a paid cloud API.
What it does not show: apart from the MTEB Code figure, the scores are in charts and the model card rather than the post, and the comparisons are Google’s own.