740M-parameter multimodal embedding model, Apache 2.0, built on Gemma 4. Natively maps text, code, images, audio, and video into one shared embedding space, all on-device. Text-only is 270M parameters with optional 170M vision and 300M audio encoders. Matryoshka Representation Learning lets you truncate vectors from 768 dims down to 128, up to 6x storage savings. Quantized, it runs in ~191MB active RAM text-only or ~567MB full multimodal on a Pixel 11 Pro. 8K token context, up from EmbeddingGemma 1 — 5.5 minutes of audio or 29 images at a go. The number that matters for practitioners: MTEB Code went from 68.76 to 78.68, a 9.92-point jump, which is the difference between "semantic code search demo" and "actually indexing your codebase locally." Weights are on Hugging Face and Kaggle, deployable via LiteRT, MediaPipe, llama.cpp, Ollama, or transformers.js, with Qdrant for vector storage and Unsloth fine-tuning guides. The full evaluation details live in the model card, so benchmark claims should be verified there before committing. Sundar presumably said "embedding" 147 times during the keynote and Google will sunset the Google AI Edge Gallery demos roughly a week before you finish reading the docs. Apache 2.0, though, so the license at least won't be deprecated. Open question for the thread: does anyone have latency numbers for the full multimodal model on mid-range hardware, or are we all extrapolating from Pixel 11 Pro figures?