Reasoning
Every model on Silk Compute thinks before it answers. You can read the thinking and choose how much of it to spend.
Thinking and effort
The thinking comes back in choices[0].message.reasoning_content and the answer in choices[0].message.content. Set reasoning_effort to low, high, or max, or to none for the least thinking the model allows.
response = client.chat.completions.create(
model="deepseek-v4.1-flash-1017-uncensored",
messages=[{"role": "user", "content": "Plan a 3-day trip to Kyoto."}],
reasoning_effort="high", # "low", "high", "max", or "none" for the least thinking
)
message = response.choices[0].message
print(getattr(message, "reasoning_content", None)) # the model's thinking
print(message.content) # the final answerLeave room in max_tokens
Thinking counts toward max_tokens. If it runs out, the reply has no text and finish_reason is length. Hy4 thinks at length at every effort, so set the limit well above the answer you expect.
With tool calls
Send the assistant message back unchanged, including reasoning_content and tool_calls, followed by the tool results. See Tool calling.
Every API format
| API | Set effort | Thinking comes back in |
|---|---|---|
| Chat Completions | reasoning_effort | message.reasoning_content |
| Responses | reasoning.effort | A reasoning output item |
| Anthropic Messages | output_config.effort | thinking blocks |
Anthropic Messages thinks only when thinking is on; without it, the model thinks as little as it can. thinking.budget_tokens works too: under 8,192 runs as low, and 32,768 or more as max.
Other effort names map to the nearest level: minimal runs as low, medium as high, and xhigh as max.