Google released a 740M multimodal embedder under Apache 2.0
Google DeepMind released EmbeddingGemma 2 on Tuesday, an open multimodal embedding model under Apache 2.0 that maps text (including code), images, video and audio into one 768-dimensional vector space. It totals 740 million parameters: a 270M text backbone with selectively loadable vision (170M) and audio (300M) encoders, so an application downloads only the modalities it uses, alongside an 8K-token context and support for more than 100 languages. It is built to run on laptops and phones rather than a hosted endpoint.
The headline feature is Matryoshka representation learning: the same embedding can be truncated to 512, 256 or 128 dimensions with minimal quality loss, which Google says cuts vector-storage costs by up to six times. On the multilingual MTEB benchmark it scores 61.36 against its predecessor’s 61.15, and Google claims a roughly 14% improvement on code tasks. The weights are on Hugging Face without a gated license, next to the model card and documentation.
WHY IT MATTERS
A 740-million-parameter embedder that runs on-device is the quiet half of the open-weights argument, and it is the half that touches the most private data: retrieval and search sit between a user and every document, message and file an app can reach, and moving them local removes the hosted API that would otherwise observe every query. The truncated embeddings are the concrete win for a builder - a six-fold cut in vector storage is the difference between a prototype and a bill - and the modular encoders mean a text-only deployment never carries the audio stack. The caveat deserves one line: the gain over the previous generation is a rounding error on MTEB, so the reason to adopt it is the footprint and the license, not the score.