The model launched on October 6, 2026, and maps all five input types into a single 768-dimensional space. Its 740 million parameters split into a 270M text model, a 170M vision encoder and a 300M audio encoder. Developers load only the encoders they need. Matryoshka Representation Learning lets vectors be truncated down to 128 dimensions, which cuts vector database storage. A single pass covers up to 5.5 minutes of audio, 29 images or 58 video frames, or a mix. The first EmbeddingGemma, released in 2025, handled text only. Google says that model passed 20 million downloads. Competing open options are larger. Alibaba's Qwen3-VL-Embedding-2B has 2B parameters and no audio support. Google's own Gemini Embedding 2 is available only as a paid API. For startups building search or RAG products, this removes the need to chain captioning, speech-to-text and text-embedding models. Google says that chaining adds latency and memory overhead. The Apache 2.0 license allows commercial use without a per-call fee. The small footprint makes private, offline retrieval on phones realistic. Teams should still test it on their own data, since the benchmark figures come from Google.