Thinking output on gemini-live-2.5-flash-preview model

[kirill-tiltgroup]

Sometimes gemini-live-2.5-flash-preview model with live connection responds with text similar to:

The user confirmed their location. This means the user is on time and there are no delays. Therefore, I should call the save_details function with the collected information. The result should be 'ON_TIME', and potential_delays should be 'None'.

Code looks like this:

   client = genai.Client(
        api_key=self.api_key,
    )
    model = "gemini-live-2.5-flash-preview"
    temperature = 0.2
    seed = 51
    top_predictions = 0.7

    config = types.LiveConnectConfig(
        response_modalities=[types.Modality.TEXT],
        output_audio_transcription=types.AudioTranscriptionConfig(),
        input_audio_transcription=types.AudioTranscriptionConfig(),
        system_instruction=self.config_manager.get_system_instruction(
            voice_name=self.voice_name,
        ),
        tools=[types.Tool(
            function_declarations=self.parse_functions(),
        )],
        thinking_config=types.ThinkingConfig(
            include_thoughts=False,
            thinking_budget=0,
        ),
        seed=seed,
        temperature=temperature,
        top_p=top_predictions,
        max_output_tokens=300,
    )`

No matter how much I try to tell in system_instruction to avoid outputting thoughts or stop thinking it doesn’t stops randomly doing that, can anybody from Google tell me how to avoid having this, or maybe if it is a bug, help solving it?

I know that there is new model for Live API gemini-2.5-flash-native-audio-preview-09-2025 but problem with using it is that in the scenario I’m trying to achieve there is only requirement to pass audio speech and receive text responses, which isn’t support by that model, and with above config I’m getting error:

Cannot extract voices from a non-audio request.; then sent 1007 (invalid frame payload data) Cannot extract voices from a non-audio request.

[llalitkumarrr]

Thank you for reaching out. Could you please check the metadata to verify if you are actually receiving thoughts? You can use the usage_metadata command for this.

Additionally just for analysis purposes, could you try providing a thinking budget for example thinking_budget=1000 and see if it works with the same model or if you encounter any errors?

[kirill-tiltgroup]

Hey @llalitkumarrr

Sure, the usage_metadata doesn't include any thoughts token information, and when I try adding the thinking_budget I got an error that this model (gemini-live-2.5-flash-preview) doesn't support this setting, but also, it seems that I found the reason for this behaviour... the system instruction that was used initially had "xml" sort of format like:

<rules>
 ....
</rules>

<message>
  ...
</message>

It was used that way as some people in internet were saying that by using this format the AI was more strict in understandings, and as soon as we reverted back to markdown format instructions, those and other tags disappeared from the responses.

On the side note, but related to latest model gemini-2.5-flash-native-audio-preview-09-2025 are there any plans to have ability to use modality TEXT for audio/text inputs but only text output?

[llalitkumarrr]

Hello @kirill-tiltgroup

This means that the output you are receiving does not include any thoughts, it is just the model’s reasoning. It seems there may be an error in the documentation indicating that thinking is supported with this model. We will analyze this further and update the documentation with the correct information.

Regarding your question about gemini-2.5-flash-native-audio-preview-09-2025, we wish to be as transparent as possible but commenting on future plans before any official release will be difficult for us. We apologize for any inconvenience this may cause.