Reasoning / :thinking

Two flavors of reasoning are exposed:

1. Native reasoning

Models that always reason — deepseek-r1, deepseek-r1-0528, deepseek-r1-distill-32b, qwq-32b. The model returns a chain of thought in choices[0].message.reasoning followed by the final answer in content.

You don’t need to ask for it — just pick the model:

curl ...chat/completions \
  -d '{"model":"deepseek-r1-0528","messages":[{"role":"user","content":"…"}]}'

2. Optional :thinking variant

Hybrid models (DeepSeek V3/V4, Kimi K2.x, GLM 5.x, Qwen 3.5+, MiniMax 2.x) have reasoning OFF by default. Append :thinking to the model id to turn it on:

curl ...chat/completions \
  -d '{"model":"nikseek:thinking","messages":[…]}'

Effort levels (only for :thinking variants):

{
  "model": "kimi-k2.6:thinking-high",
  "messages": [...]
}

Response shape

For reasoning calls, the gateway adds a nikcli.thinking block:

{
  "choices": [{
    "message": {
      "role": "assistant",
      "content": "9",
      "reasoning": "Need to compute 2+7 = 9. So the answer is 9."
    }
  }],
  "nikcli": {
    "thinking": { "requested": true, "native": false, "effort": "medium" },
    "resolvedModel": "deepseek-v4-pro",
    ...
  }
}

Which models support what

See Models — the :thinking column shows which are hybrid.