EmbeddingGemma 2: An open, lightweight multimodal embedding model

	EmbeddingGemma 2: An open, lightweight multimodal embedding model(blog.google)
	417 points by ilreb 7 days ago \| 46 comments
	tl;dr: Google released EmbeddingGemma 2, a 740M-parameter multimodal embedding model that maps text, code, images, audio, and video into a unified space for on-device use. Built on Gemma 4 and licensed under Apache 2.0, it features a modular design (270M text-only, with optional 170M vision and 300M audio encoders), an 8K context window, and Matryoshka Representation Learning for up to 6x storage reduction. It claims best-in-class performance for sub-1B models on MTEB Code and MAEB, running in ~191MB RAM for text-only on a Pixel 11 Pro.
	HN Discussion: ↑Apache 2.0 licensing is essential for embedding models since vendors deprecating proprietary models breaks stored vectors ↑Lightweight on-device multimodal embeddings fill an important gap for local/private use cases ↑Appreciation that Google open-sourced a model likely close to what ships in their own products ~Technical limitation noted: uses MRL rather than MatFormers, so model weights can't be shrunk alongside embeddings •Questions about comparisons to alternatives like Voyage AI and compatibility with binary quantization approaches