The local-RAG default just moved. Google’s EmbeddingGemma 2 is a 744M-parameter embedding model, Apache 2.0, that turns text (and, per the family line, image and audio inputs) into a vector a search index can score. 21,148 direct downloads and 1,087 likes in days; community builds (unsloth GGUF at 29,700 downloads, an ONNX export, a LiteRT-LLM port for phone GPUs) landed within the week.
Why embeddings matter more than chat models do locally. A retrieval stack runs the embedding model on EVERY query and EVERY document - thousands of calls where a chat model runs one. At 744M params the fp16 build is about 1.5GB and a 4-bit one is about 0.4GB, so the whole index-brain fits in RAM that would otherwise hold one layer of a chat model, and the 48MB-per-1,000-token KV cost barely registers. That is the difference between an index that re-embeds weekly and one that re-embeds nightly.
Where it sits. Google’s original embeddinggemma (300M) proved phone-scale retrieval at 3.6M downloads; version 2 brings the 8k-context class, and the open-weights Apache license puts it next to BGE-M3 and GTE as the default pick for a self-hosted RAG stack that wants a current-generation embedding model without a server.