Parameter Compatibility

Which parameters work with which models.

Guaranteed Compatible (all models)

Parameter	Description
`model`	Model name
`messages`	Message array
`stream`	Enable SSE streaming
`max_tokens` / `max_completion_tokens`	Max output tokens
`temperature`	Sampling temperature
`top_p`	Nucleus sampling
`stop`	Stop sequences
`seed`	Deterministic sampling

Widely Supported (most models)

Parameter	Description
`tools` / `tool_choice`	Function calling
`response_format`	JSON mode / structured outputs
`frequency_penalty`	Repetition reduction
`presence_penalty`	Topic diversity
`logprobs` / `top_logprobs`	Token probabilities

Model-Specific

Parameter	Models
`top_k`	Anthropic, some open models
`repetition_penalty`	Open models (Llama, Mistral)
`min_p`	Some open models

How It Works

The marketplace passes through all parameters to the provider unchanged. Unsupported params are silently ignored by most providers (same behavior as OpenRouter).

The stream_options: {"include_usage": true} param is auto-injected for metering — you don't need to add it.