Google has released an update to the Gemma 4 model that modifies its chat templates. The changes address significant issues with tool calling capabilities and reduce "laziness" in the model's responses.

  • Fixes major bugs in tool calling functionality.
  • Reduces "laziness" in output generation.
  • Enables Flash Attention 4 support on Hopper GPUs.
  • Provides an interactive guide for improving vision capabilities.

These updates aim to improve the reliability and efficiency of Gemma 4 for developers and users.