Gemini 4 Breaks the Mold: Google DeepMind Ships Early Post-Training Model

Google DeepMind has announced that its next flagship model, Gemini 4, has entered post-training. However, unlike previous releases, Gemini 4 will be shipped as an early post-training checkpoint rather than a fully polished release.

Koray Kavukcuoglu, Google DeepMind's senior vice president, stated during the AI Agenda Live Summit that the company intends to roll out an early post-training version 'as soon as possible.' This decision marks a departure from Google's previous approach of releasing models after extensive safety fine-tuning and testing.

The move is seen as a trade-off between speed and polish. Shipping an early post-training model allows for faster release, but it may also lead to unpredictable behavior and require more frequent patches after launch.

This decision puts Google in line with other labs like OpenAI, which has been iterating publicly on its models for years. However, the move also raises concerns about safety and stability, as early post-training models have had less safety fine-tuning and testing.