Parameters and model support

Parameter support and limits vary by model. The catalog is the source of truth for each model.

The model catalog lists supported_parameters and passthrough_parameters for each model.

Common parameters

Parameter	Purpose
`max_tokens`	Set the maximum output length.
`temperature`	Change output variation.
`stream`	Return server-sent events.
`tools`	Give the model tool definitions.
`service_tier`	Select `fast`, `default`, or `flex`.

Gradiated rejects unsupported parameters. It does not remove them silently. The response uses the code unsupported_parameter and names the parameter when possible.

Context limits

The request must fit the selected model context. This includes messages, tools, and the requested output length.

Reduce the input or max_tokens when the API returns context_length_exceeded.