Gemma 4 Trips Over Chat Templates

This looks less like a model meltdown and more like a brittle local-inference stack exposing how sensitive Gemma 4 is to formatting details.

• –Several commenters point to Ollama, llama.cpp, and quant/template mismatches as the likely cause, which matches the pattern of “works in one wrapper, breaks in another.”
• –Gemma 4’s release leans hard into tool calling, structured output, and thinking modes, so older templates or partial support can degrade behavior fast.
• –The lesson for local users is simple: when a new open model ships, update the runtime first, then blame the weights.
• –The post is still useful signal because it shows the model can fail in confusing ways when prompt formatting is off, especially for reasoning-heavy workflows.