People are finding really creative ways to use EmbeddingGemma 2 beyond search. Instead of running a full multimodal LLM, this community project trains a lightweight classifier on top of EmbeddingGemma 2 in real time. Because the model is compact and natively multimodal, it can learn new categories on the... fly and classify inputs locally in ~20ms for text and ~100ms for images.show more

Project by @mmastrac! He also open-sourced the approach as ‘hunch’ if you want to try it yourself: Demo post: GitHub:

EmbeddingGemma 2 is proving that you don't need a massive multimodal LLM for every task.

Local classification in milliseconds without a huge model is brilliant

I am in love with embedding Gemma 2.

Can a category taught with text examples transfer to images in hunch, or does each modality need its own labels? With EmbeddingGemma 2, that could make “teach once, sort both documents and photos” a useful product pattern.

This is great for app builders who need cheap on-device sorting, like tagging photos or flagging support tickets. I'd use EmbeddingGemma 2 plus a tiny classifier instead of calling a big model per item, since it costs far less and keeps data on the phone.

This guy should have studied at MIT 😐😅

Hey @googlegemma is it possible to get a shout out for this project - using EmbeddingGemma model on a @arduino VentunoQ for multi-modal embedding for clothing suggestions 😁

Detail from the hunch repo: it runs on plain CPUs and only takes over once it agrees with its teacher model often enough, then keeps getting spot checked. Cheap reflexes for agents 🧠

embedding 顶上套个轻量分类器这招太实用了，实时训练加端侧推理，成本直接砍掉一个量级。好奇问下，实时训练的标注是用户现场给的吗？冷启动时类别很少的情况下效果怎么样

how small is the classifier trained on top of embeddinggemma?

Carbon - where is it?

Freezing the embedding and only training the head is the real win here — but real-time retraining means your classifier drifts with every new label batch. Version the embedding index or you'll chase ghost regressions.

how many examples does a new category need before it holds up? if it's under 10, a support team could train its own ticket router without ever calling an ml person

a head trained live is only as good as its handful of labeled examples, and drift shows up as inputs landing between categories

Yup nice

Amazing

embedding models quietly became the most useful part of the stack. everyone's chasing bigger llms while the real production wins keep coming from small, fast, multimodal embeddings doing 80% of the job at 1% of the cost