Gemma4 llm node

[der]

gemma4:26b is very good and fast used directly but from pydantic within the ROS llm node it is stuck in reasoning mode which makes it very slow to respond and not useful for front line Marvin responses.

Can we find a solution to the thinking config?

Alternative is to use a lightweight model for immediate responses and use gemma as an background agent that gets triggered for hard tasks.

[der]

Set explicit ModelProfile to declare that it supports reasoning seems to have helped.

        ollama_model = OpenAIChatModel(
            model_name=self.model_name,
            provider=OllamaProvider(base_url='http://localhost:11434/v1'),
            profile=ModelProfile(
                supports_thinking=True,
                thinking_tags=('<|channel>thought','<channel|>')
            )
        )

        self.agent = Agent(
            ollama_model,
            output_type=str,
            system_prompt=(
                'You are small droid called Marvin with speech, vision and movement capabilities.'
                'Respond to questions VERY BRIEFLY in plain text that the droid can speak aloud.'
                'If the user just says "Marvin" then respond with "Hi"'
            ),
            model_settings={'thinking': False}
        )