maxOutputTokens: the default (8000), but not more than the model's
max output, and not more than a quarter of its context window.
contextBudgetTokens: the default (32000), but not more than 75% of
what is left of the window after the output.
the steps, the tool result cap and the temperature stay the default.
Then override (the ai.json override of the model) may set
chatMaxSteps, contextBudgetTokens and maxOutputTokens as they are.
Override fields that are not positive integers are ignored.
Gemma 4 (262k window, 16k max output) gets exactly DEFAULT_BUDGETS.
The budgets of a chat with a model:
maxOutputTokens: the default (8000), but not more than the model's max output, and not more than a quarter of its context window.contextBudgetTokens: the default (32000), but not more than 75% of what is left of the window after the output.Then
override(theai.jsonoverride of the model) may setchatMaxSteps,contextBudgetTokensandmaxOutputTokensas they are. Override fields that are not positive integers are ignored.Gemma 4 (262k window, 16k max output) gets exactly DEFAULT_BUDGETS.