On October 6, 2026, Google released EmbeddingGemma 2, which it describes as "the most capable model for on-device multimodal embeddings." The first EmbeddingGemma (September 2025) handled text only; SiliconANGLE reports Google DeepMind's engineers counted more than 20 million downloads of it. Weights are on Hugging Face and Kaggle under Apache 2.0.
What changed technically
- One space for five kinds of input. Text, code, images (including PDFs, slides and charts), video frames and raw audio all map into the same 768-dimensional space, so a spoken query can be compared directly with a video clip.
- Load only what you use. The developer guide lists four configurations from one checkpoint: 270M (text/code), 440M (text + vision), 570M (text + audio) and 740M (full multimodal). Google says the quantized text-only weights need as little as ~191MB of active RAM on a Pixel 11 Pro, and the full model ~567MB.
- Longer inputs. An 8K-token context, four times the first version, enough for "up to 5.5 minutes of audio, 29 images, 58 video frames."
- Cheaper vectors. Matryoshka Representation Learning lets developers cut vectors to 512, 256 or 128 dimensions. Google's guide says that at 256 dimensions image, video and speech retrieval keep "about 95%" of full quality, and that a million vectors drop from roughly 1.5 GB to 250 MB at 128 dimensions.
What developers are building with it
The largest benchmark gain is on code: MTEB Code rises from 68.76 to 78.68, and Google pitches the model at "local codebase indexing, semantic code search, and coding agent retrieval." Because it shares Gemma 4's text tokenizer and audio encoder, an on-device RAG pipeline running both needs less memory than two unrelated models. Google's own demos include a Video Moments Finder that locates a scene from a typed or spoken query, and a Mac meeting app, AI Edge Foresight, that pairs the two models. Serving support at launch spans sentence-transformers, vLLM, llama.cpp, Ollama, MLX and transformers.js, with Gemini Enterprise Agent Platform availability "coming soon."
Why it matters. Retrieval for agents has mostly meant calling a hosted embedding API, which sends every indexed document off the device. A small, permissively licensed model that indexes code, screenshots and recordings locally changes the privacy and latency math for desktop agents and phone assistants, and the modular loading means a text-only coding tool doesn't pay for audio it never uses.
What remains uncertain. The headline claims -- best quality-per-parameter under 1B, beating "some specialist models more than twice its size" -- are Google's, from its own model card. The ~95% retention figure applies to media retrieval at 256 dimensions. Analysis: teams should measure truncation on their own corpus before shrinking an index.
EmbeddingGemma 2, released October 6 under Apache 2.0, gives on-device apps and coding agents one embedding space for text, code, images, audio and video in 270M-740M parameters, with a near-10-point code-retrieval gain and up to 6x smaller vectors -- though its quality claims are Google's own benchmarks.