EmbeddingGemma 2

EmbeddingGemma 2 is an open-weight embedding model from Google, released on 6 October 2026 under the Apache 2.0 licence with weights on Hugging Face and Kaggle; availability in Google’s enterprise Model Garden is described as coming soon. Built on the Gemma 4 architecture, it maps code, text, images, audio and video into one shared embedding space, so a search can match a voice memo to a video clip or a text query to an audio recording. Google gives 740 million parameters for the full multimodal model, of which a 270 million parameter core handles text alone, with optional vision (170M) and audio (300M) encoders. It accepts an 8K token context and produces 768-dimension vectors that can be truncated to 512, 256 or 128 dimensions through Matryoshka representation learning, which Google says cuts vector storage by up to six times. Google reports about 191MB of active memory for the text-only weights on a phone, which makes it suited to on-device and private retrieval where documents should not leave the device. The benchmark comparisons in Google’s post are its own, so no score is stored here. There is no hosted price, since the model is downloaded and run by the user.

Not yet independently verified: Only Google’s own announcement of 6 October 2026 has been read; no independent dated report is cited yet. The benchmark comparisons in the post are Google’s own, so no score is stored. We will update this when it can be confirmed, and remove this note.