Reasoning effort

Reasoning effort controls how much work a supported model does before producing its answer. Higher effort can improve difficult reasoning tasks, but may increase latency and output-token usage.

Doubleword accepts the OpenAI reasoning effort levels: none, minimal, low, medium, high, xhigh, and max.

Support varies by model. If a model cannot represent an effort, the request is rejected rather than silently changed to a different level.

Model support

Reasoning capability data is not currently available.

Chat Completions

Set reasoning_effort on a Chat Completions request:

curl https://api.doubleword.ai/v1/chat/completions \
  -H "Authorization: Bearer {{apiKey}}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "{{selectedModel.id}}",
    "messages": [
      {
        "role": "user",
        "content": "Solve this problem carefully and explain the answer."
      }
    ],
    "reasoning_effort": "medium",
    "max_completion_tokens": 16384
  }'

The legacy max_tokens field can be used instead of max_completion_tokens.

Qwen3 non-thinking mode

Some Qwen3 deployments do not advertise none as a reasoning effort. To request the native non-thinking mode, append /no_think to the latest user prompt:

{
  "model": "Qwen/Qwen3-14B-FP8",
  "messages": [{
    "role": "user",
    "content": "Calculate 17 × 19. Return only the final integer. /no_think"
  }]
}

/no_think is Qwen-specific prompt syntax, not an OpenAI reasoning-effort value. An empty reasoning block may still be returned for API compatibility. See the Qwen3 switching documentation.

Responses

Set reasoning.effort on a Responses API request:

curl https://api.doubleword.ai/v1/responses \
  -H "Authorization: Bearer {{apiKey}}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "{{selectedModel.id}}",
    "input": "Solve this problem carefully and explain the answer.",
    "reasoning": {
      "effort": "medium"
    },
    "max_output_tokens": 16384
  }'

Notes

• Omit the reasoning field to use the model's default behavior.
• none requests a non-reasoning response, but reasoning-only models may reject it.
• Use only the OpenAI-compatible fields above. Do not send provider-specific fields such as thinking, thinking_token_budget, or chat_template_kwargs.
• Some models map efforts to an absolute reasoning-token budget. In that case, the output limit must be greater than the mapped budget. Doubleword rejects an insufficient limit rather than clipping the reasoning budget.