EmbeddingGemma 2

EmbeddingGemma 2 at a glance

Google DeepMind released EmbeddingGemma 2 on October 6, 2026. The 740-million-parameter model maps text and code, images, video and audio into a shared 768-dimensional embedding space. Its weights are available on Hugging Face under an Apache 2.0 license. The model card describes a 270-million-parameter text component plus optional vision and audio encoders, so developers can load only the modalities they need.

Deployment and limits

Google positions the model for local search, retrieval and classification on consumer hardware. It supports an 8,192-token context and output vectors that can be shortened to 512, 256 or 128 dimensions, with a quality tradeoff. The model card warns that 128-dimensional vectors substantially reduce multimodal quality and that float16 inference can produce invalid or degraded embeddings. Google’s benchmark comparisons are its own evaluations; developers should test accuracy, memory use and language coverage on their own data before deployment.