FAQ : Is their a way to turn off thinking mode ? (disable thinking system prompt obviously not work)

[HuangChiEn]

Although thinking mode only support in 27b version as doc, it lead the debug become a bit hard.
As title said, is their anyway to disable the thinking mode ?

Of course disable thinking system prompt doesn't work, seems model will enable thinking structure by defualt.

[KarasuZero]

I'm running into the same issue, have you found a solution for this problem? in my case I need to the model to output info in a specific format, having thinking chunks racks up a significant ammount of processing time

[HuangChiEn]

I'm running into the same issue, have you found a solution for this problem? in my case I need to the model to output info in a specific format, having thinking chunks racks up a significant ammount of processing time

no, disable prompt doesn't help. It depend on the problem. If the problem need further thinking, thinking mode is still enabled.
But give the system prompt i.e. "Directly answer the user question without (over)thinking" may work in some degree, it'll also downgrade the performance thought.

[srikantak-221]

Hi @HuangChiEn ,

The “Thinking” behaviour in MedGemma 27B is not a separate module or runtime feature that can be disabled; it is a behavioural byproduct of the model’s fine-tuning. Because MedGemma is optimized for medical accuracy and reduced hallucination risk, it is trained to surface intermediate reasoningas part of its response style. There is no inference-time flag to turn this off. Please refer to this colab: https://github.com/google-health/medgemma/blob/main/notebooks/quick_start_with_model_garden.ipynb The Colab example controls this behavior through prompt engineering rather than a model switch. If you need to suppress visible reasoning, for debugging or production, the supported approach is to explicitly instruct the model to provide only the final concise answer and to constrain generation parameters such as max tokens and temperature. The model will still perform internal reasoning, but prompt-level guidance determines whether that reasoning is exposed in the output.

Thank you!