Google describes a real-time classifier built on EmbeddingGemma 2

Google's Gemma account says a community project trains a lightweight classifier on EmbeddingGemma 2 in real time, instead of running a full multimodal LLM. The post says the classifier can learn new categories on the fly and runs locally in about 20ms for text and 100ms for images.

The project's author, mmastrac, has open-sourced the approach as hunch, which exposes it as a real-time /v1/systemone endpoint. The author says it is multimodal by default and can distill an upstream teacher model.