Model integration · Local message analysis

Why keep AI on the phone to detect suspicious chats?

GovTech Singapore and Google DeepMind adapt Gemma for local scam detection. Lightweight screening precedes conversation analysis: a small model meets an everyday need.

Announced October 14, the model will reach applications in phases; a universal activation route has not been announced.

In this articleContents3

中文

Screen first, then examine the conversation

Local scam conversations inform the adapted Gemma. A low-power classifier screens messages first; flagged cases reach Gemma for longer-context analysis and a possible pause warning. Reserving deeper analysis for those cases is the design's response to battery costs.

Official abstract artwork with red-and-white layers, blocks and glass spheres.
Official announcement artwork; it does not show product operation. · Open full-size image
A lightweight classifier screens messages; flagged conversations reach Gemma, which can trigger a warning. Both stages can return a benign verdict.
BitShovel bilingual diagram of the official flow: both stages run locally. It is not an accuracy or battery measurement. · Open full-size image
Sources and further reading

中文

A small model's value also comes from where it is used

Gemma provides models that developers can adapt and integrate. Our reading is that this project connects local context, a classifier and a phone warning. Local notes and EmbeddingGemma retrieval offer related examples of moving a capability closer to material, while solving different tasks with separate evidence.

Sources and further reading