How EmbeddingGemma 2 Finds a Moment From a Voice Memo
Summary
EmbeddingGemma 2 is Google DeepMind's open embedding model that puts text, images, video, and audio into one space, built for on-device use. Video is sampled at one frame per second and each frame becomes 140 tokens; audio becomes 25 tokens a second. Every input runs through one shared 24-layer backbone and is mean-pooled into a single 768-number point, so a spoken query lands next to the clip it describes and search becomes a nearest-point lookup. Matryoshka truncation keeps the first 256 numbers for an index three times smaller, with benchmark scores down by three points or less. The full model is 740M parameters, and you can load only the encoders you need (text alone is 270M). Model page: deepmind.google/models/gemma/embeddinggemma
About this video
EmbeddingGemma 2 is Google DeepMind's open embedding model that puts text, images, video, and audio into one space, built for on-device use. Video is sampled at one frame per second and each frame becomes 140 tokens; audio becomes 25 tokens a second. Every input runs through one shared 24-layer backbone and is mean-pooled into a single 768-number point, so a spoken query lands next to the clip it describes and search becomes a nearest-point lookup. Matryoshka truncation keeps the first 256 numbers for an index three times smaller, with benchmark scores down by three points or less. The full model is 740M parameters, and you can load only the encoders you need (text alone is 270M). Model page: deepmind.google/models/gemma/embeddinggemma