Google releases EmbeddingGemma 2: an open 740-million-parameter model that understands text, code, images, video and audio
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
Multimodal search across text, video and audio is becoming light enough to fit in a phone, and fully open at that.
The model encodes all five data types into a single space, and for text-only work it can be trimmed to 270 million parameters. Its context is 8 thousand tokens (up to 5.5 minutes of audio, 29 images or 58 video frames); on a Pixel 11 Pro it needs about 567 MB of memory. According to Google, on the MTEB Code benchmark the model scores 78.68 points (9.92 higher than the first version) and on image and audio tasks beats some specialized models twice its size. The license is Apache 2.0, with weights on Hugging Face and Kaggle.
Sources 2