Reasoning
Read the model's reasoning and allow enough output tokens for its answer.
raon-1-mini-preview reasons before it answers. Each API returns the reasoning apart from the answer:
| API | Where the reasoning is |
|---|---|
| Chat Completions | message.reasoning (delta.reasoning when streaming) |
| Responses | a reasoning item in output |
| Messages | a thinking content block |
Reasoning and max_tokens
Reasoning counts toward the output limit (max_tokens, max_completion_tokens or max_output_tokens). If the limit is used up while the model is still reasoning, the answer can be empty. Chat Completions reports finish_reason: "length", Messages reports stop_reason: "max_tokens", and Responses reports status: "incomplete" with incomplete_details. Leave room for both reasoning and the answer, and increase the output limit if needed.
Always enabled
Raon always reasons before answering. Reasoning cannot be disabled. The API accepts chat_template_kwargs.enable_thinking for compatibility but ignores its value, including false.
Reasoning effort
Reasoning cannot be disabled and its effort or token budget cannot be tuned. These parameters are accepted without an error but do not change the result:
| API | Accepted, no effect |
|---|---|
| Chat Completions | reasoning_effort |
| Responses | reasoning.effort |
| Messages | thinking (including type: "disabled" and budget_tokens) |