Use the recommended sampling settings
prompting & control tokens
// what it does
Google recommends temperature 1.0, top_p 0.95, and top_k 64 for Gemma 4 across use cases. In the Ollama REPL, /set parameter changes them live; in API calls pass them in the request body; llama.cpp uses --temp, --top-p, and --top-k. Lower the temperature for more deterministic extraction or classification.
// shell
$ /set parameter temperature 1.0$ /set parameter top_p 0.95$ /set parameter top_k 64// gotcha
Gemma 4's recommended temperature of 1.0 is higher than many models' 0.7 default. Reusing a 0.2-0.4 preset copied from another model can make Gemma 4 feel oddly flat or repetitive — start from the recommended values.