Google has launched the gemma-trainer skill, allowing users to fine-tune custom Gemma models locally. This tool shifts computing power and data resources towards local devices, with revenue generated through supporting hardware and enterprise services, motivated by expanding the influence of the Gemma ecosystem and collecting feedback from distributed training to optimize the base model. The current open-source large model industry is in a transition phase from pre-training to efficient fine-tuning and local deployment, lowering barriers to accelerate adoption.