Gemma 4 improves tool use and adds a vision token guide
Google is revising the settings that format conversations for Gemma 4. The update includes major fixes for and aims to reduce “laziness,” where the model stops early or does less than requested. Support for preserve_thinking is also being added.
will be enabled on Nvidia’s Hopper GPU generation. An interactive guide will help people test and improve the model’s vision features, with a linked space focused on the vision .
Key points
- Google is updating Gemma 4’s conversation-formatting settings.
- The changes target failures and answers that do too little.
- Support for preserve_thinking is being added.
- will work on the Hopper GPU generation.
- An interactive tool covers vision features and their .