Fine-Tuning Small Models into Usable On-Device Agents

joemeno · x · 2026-07-17

Google engineers shared a pipeline for turning small models into usable offline agents directly on mobile phones: starting with Gemma 270M, fine-tuning with synthetic task data, followed by LoRA training and int4 quantization, before finally deploying for on-device execution on Pixel.

The author emphasizes that the goal here isn't to chase a "general-purpose LLM," but rather to enable a 270M small model to outperform much larger 70B models on specific tasks, while still achieving extremely high inference speeds in an on-device environment. The post summarizes this stack as: Gemma 270M + synthetic data + LoRA + int4 quantization + on-device runtime.

Original post →

More from coding & agent

coding & agent channel →