EmbeddingGemma 2

Changes from previous version

First EmbeddingGemma version in this catalog. Adds code, image, video, and audio to the text embeddings of the prior EmbeddingGemma release, with a 4x context window and a 9.92-point gain on MTEB Code.

Release Summary

Open Apache 2.0 embedding model built on Gemma 4. 740 million parameters in full, or 270 million for text and code, plus optional vision (170 million) and audio (300 million) encoders. It maps text, code, images, video, and audio into one 768-dimensional vector, which Matryoshka truncation can cut to 512, 256, or 128. Context is 8K tokens, four times EmbeddingGemma, enough for about 5.5 minutes of audio, 29 images, or 58 video frames. Google reports 78.68 on MTEB Code, 9.92 points above EmbeddingGemma at 68.76. Quantized text weights use about 191MB of RAM on a Pixel 11 Pro, and the full multimodal model about 567MB.

Timeline