Turn the reasoning channel off on Gemma 4, in every request

Practice

Turn the reasoning channel off on Gemma 4, in every request

Status	current
Confidence	high
Applies to	Gemma 4 26B A4B llama.cpp and llama-server
Source	mission-control GPU-LLM.md, rule R1 2026-08-25
Updated	2026-10-09

The practice

Send chat_template_kwargs {enable_thinking: false} with every call, even though the server defaults it off. A grammar-constrained call went from 600 tokens and 8.7 s to 72 tokens and 1.4 s. The server default exists for clients that cannot send the field, not as permission to skip it.

2026-10-09

Gemma 4 emits 150 to 500 tokens of hidden chain-of-thought before every answer, returned in a separate reasoning field. Only the enable_thinking template argument turns it off; reasoning_budget: 0 and reasoning_effort: none are accepted and do nothing. Keep sending the field because nothing else on the system has the server default, and a model swap or a drop-in rollback would silently bring the thinking back for anyone relying on it.

Comments

Public comments on each entry are coming. Nothing is collected here yet.