Developers / Docs / Reasoning

Reasoning

Reasoning models can work through a problem before they answer. On ARK, thinking is off unless you ask for it, so simple requests stay fast and cheap.

Which models reason

GET /api/v1/models marks reasoning models with "ark_reasoning": true. On ARK Cloud these are arklabsai/Qwen3.6-35B-A3B, arklabsai/Qwen3.5-122B-A10B, arklabsai/Qwen3.8-Flash-Next and arklabsai/Qwen3.8-27B.

Turn thinking on

Set reasoning_effort. ARK accepts OpenAI's whole range of values and maps them onto three levels:

You sendLevelEffect
nothing, none, minimal, lowlowNo thinking. The model answers directly.
mediummediumThinking on.
high, xhigh, maxhighThinking on.

The Qwen models think the same way at medium and high. reasoning: {"effort": "high"} is accepted as an alternative spelling.

response = client.chat.completions.create(
    model="arklabsai/Qwen3.5-122B-A10B",
    messages=[{"role": "user", "content": "A train leaves Szczecin at 9:12 and arrives in Warsaw at 14:47. How long is the trip?"}],
    reasoning_effort="high",
    max_tokens=4096,
)
message = response.choices[0].message
print("thinking:", message.reasoning)   # the SDK keeps unknown fields
print("answer:", message.content)
print("reasoning tokens:", response.usage.completion_tokens_details.reasoning_tokens)

Where the thinking goes

  • Non-streamed: the thinking is in choices[0].message.reasoning and the answer in content. The field is called reasoning, not reasoning_content.
  • Streamed: thinking arrives first as delta.reasoning, then the answer as delta.content.
  • Usage: completion_tokens_details.reasoning_tokens counts the thinking. Those tokens are part of completion_tokens and are billed as output.
Leave room for thinking

max_tokens covers thinking and answer together. A hard problem can spend thousands of tokens thinking; if the limit is hit first, finish_reason is length and the answer may be missing. Give reasoning requests a generous limit.

Reasoning and tools

Thinking works with tools. With reasoning_effort medium or high, the Qwen models think first and then call a tool or answer, streamed or not. At low, the default, they call tools without thinking.

Thinking in the conversation history

You do not need to send earlier thinking back. A reasoning field on an assistant message in messages is accepted and dropped.

With the Anthropic API

On /api/anthropic/v1/messages, use thinking: {"type": "enabled", "budget_tokens": 8192} or more selects the high level, a smaller budget the medium level, and {"type": "adaptive"} the medium level. The thinking comes back as a thinking content block. budget_tokens picks the level; it is not a hard cap. See Messages API.