An experiment with the Gemini API revealed that setting a `thinkingBudget` to 0 did not prevent the model from using thinking tokens, contrary to documentation for Firebase AI Logic. The API accepted requests with a `thinkingBudget` of 0, 1024, and 8192, and the `gemini-4-hydrogen` model consistently used over 1,200 thinking tokens even with a budget of 0. This suggests that the `thinkingBudget` parameter acts as a suggestion rather than a strict limit for this model and API. AI
IMPACT Developers should verify LLM configurations and not assume parameters like `thinkingBudget` are strict limits without explicit documentation.
RANK_REASON The item describes an unexpected behavior of an API parameter, which is a specific product observation rather than a new model release or significant industry event.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →