Gemma 4 fails multi-turn tool calling

Gemma 4's native function calling is a welcome addition, but its brittle execution in local environments highlights the gap between hosted APIs and open-weight models.

• –The bug triggers specifically during extended agentic loops (think -> tool -> respond -> think -> tool), causing an immediate generation halt.
• –This severely limits the model's viability for autonomous workflows that require continuous, multi-step reasoning and interaction.
• –The failure likely stems from how local inference engines or quantization formats parse the model's new thinking and tool-use tokens.
• –Fixes will likely require client-side updates in inference frameworks like llama.cpp or Ollama to properly handle the token streams.