Gemma 4 Chat Template (llama.cpp / OpenWebUI)

Custom Gemma 4 chat template designed for use with llama.cpp and OpenWebUI.

Overview

This template is based on the newer official Gemma 4 format, but modified to fix an issue where reasoning-channel tokens such as thought and <|channel> could leak into visible outputs.

The goal is to:

• Keep modern tool-calling behavior
• Maintain compatibility with llama.cpp
• Prevent thinking-channel leakage in responses

Example

Before (broken output)

After (fixed output)

Changes from the Official Template

• Removed replay of:
  • message.reasoning
  • message.reasoning_content
• Removed forced empty:
  • <|channel>thought ... <channel|>
• Kept:
  • tool-call handling
  • tool-response formatting
  • assistant continuation logic

Why This Exists

The official Gemma 4 template assumes a serving stack that properly handles reasoning channels.

In some setups, such as llama.cpp with OpenWebUI, those markers may appear in the final output. This template removes the problematic parts while keeping the newer tool and message flow intact.

Usage

Replace your current chat template with the modified version.

Tested with:

• llama.cpp (peg-gemma4)
• OpenWebUI
• Gemma 4 26B (Bartowski GGUF)

Notes

You may still see this warning in llama.cpp:

common_chat_try_specialized_template: detected an outdated gemma4 chat template, applying compatibility workarounds. Consider updating to the official template.