Google researchers introduce Gemini for Google, a custom LLM adapted for internal software engineering that reduced developer iteration counts by 23% in a blind A/B study across 29,000 developers.

Google developed Gemini for Google (GfG), a specialized version of Gemini fine-tuned on a trillion-token proprietary dataset of internal software engineering data, with a mid-training strategy designed to prevent catastrophic forgetting (loss of previously learned capabilities).

In a large-scale blind A/B study across 29,000 developers, the model reduced the mean number of iterations per turn by 23% and increased code survival rates by about 17%, compared to baseline LLMs.

The paper provides a blueprint for enterprise model adaptation covering: extraction of high-value signals from software engineering data, data preparation strategies, full-stack model tuning (continued pre-training and post-training), and deployment of downstream applications.