The developer of the Circles app created a custom, on-device LLM to process dictated notes into structured contact information, avoiding cloud-based processing for privacy and cost reasons. Initial attempts with a general-purpose Gemma 4 E2B model proved too slow and less accurate. By fine-tuning a much smaller Gemma 3 270M model on a synthetic dataset of 18,000 notes, the team achieved a significant improvement in both speed and accuracy, processing notes in about one second on a phone's CPU. AI
IMPACT Demonstrates the viability of highly specialized, on-device LLMs for privacy-sensitive data processing in consumer applications.
RANK_REASON This is a description of a specific application of an LLM for a particular software product, not a release of a new model or research.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
