reasoning_effort to select a model’s reasoning level and reasoning_format to control how the API returns its reasoning.
Reasoning tokens count toward max_completion_tokens and the completion-token usage reported by the API. Effort levels select model-specific reasoning modes; they don’t reserve or guarantee an exact token budget.
Supported models
The following models support reasoning. Select a model to see its supported effort levels and an example request.
For Shared Inference, use the model ID listed in the model catalog. For Dedicated Inference, use the endpoint ID assigned to your organization as the request’s
model value.
Send a reasoning request
Install the Cerebras Python SDK and setCEREBRAS_API_KEY as described in the Quickstart. This example selects medium effort for GPT OSS 120B and prints the reasoning separately from the final answer.
choices[0].delta.reasoning and final-answer text from choices[0].delta.content.
The remaining Python examples reuse client. In Dedicated Inference examples, replace YOUR_ENDPOINT_ID with your assigned endpoint ID for that model.
Reasoning format
Usereasoning_format to select how the API returns reasoning. Support varies by model.
The following table summarizes format support. A value of
none for reasoning_format doesn’t disable reasoning. To disable reasoning on a model that supports it, use reasoning_effort="none".
Model behavior
Each section lists the model’s reasoning effort values, shows an example request, and describes its format restrictions.Qwen 3.8 27B
qwen-3.8-27b enables reasoning by default. Use reasoning_effort to select one of the following modes.
This example selects
high effort. To disable reasoning, set reasoning_effort to none.
message.reasoning or streaming delta.reasoning. Setting reasoning_format to raw doesn’t change this behavior. The hidden format isn’t supported.
Conversation history: Use clear_thinking to preserve or remove historical assistant reasoning. See Multi-turn reasoning.
Do not send Qwen-native parameters such as disable_reasoning, enable_thinking, preserve_thinking, or thinking_budget. Use reasoning_effort and clear_thinking instead.
Kimi K2.7 Code
kimi-k2.7-code always enables reasoning. The API accepts reasoning_effort for client compatibility but ignores its value.
This example requests parsed reasoning and omits
reasoning_effort because it has no effect.
parsed. The raw format is supported; hidden isn’t supported.
Conversation history: clear_thinking isn’t implemented. Do not send it.
Kimi K3
kimi-k3 always enables reasoning. Use reasoning_effort to select one of the following levels.
This example selects
high effort and parsed reasoning.
parsed is supported. Omit reasoning_format or set it to parsed. Do not send raw or hidden.
Conversation history: Replay the complete assistant response, including reasoning and tool calls. Kimi K3 preserves reasoning history. Do not send clear_thinking.
Kimi K3.1 Pro
kimi-k3.1-pro always enables reasoning. Use reasoning_effort to select one of the following levels.
This example selects
medium effort and parsed reasoning.
parsed is supported. Omit reasoning_format or set it to parsed. Do not send raw or hidden.
Conversation history: Replay the complete assistant response, including reasoning and tool calls. Kimi K3.1 Pro preserves reasoning history. Do not send clear_thinking.
GLM 5.3 Flash
zai-glm-5.3-flash always enables reasoning. Use reasoning_effort to select one of the following levels.
This example selects
high effort and parsed reasoning.
parsed. Both raw and hidden are supported. The hidden format omits reasoning text while the model still generates and counts reasoning tokens.
Conversation history: Use clear_thinking to preserve historical assistant reasoning or remove reasoning before the latest user message. See Multi-turn reasoning.
GPT OSS 120B
gpt-oss-120b always enables reasoning. Use reasoning_effort to select one of the following levels.
This example selects
medium effort and parsed reasoning.
parsed. Both raw and hidden are supported. With raw, reasoning and final-answer content are concatenated without separators.
Gemma 4 31B
gemma-4-31b disables reasoning by default. Use reasoning_effort to enable or disable reasoning.
This example enables reasoning with
medium effort and requests parsed reasoning.
raw and hidden formats aren’t supported.
Conversation history: clear_thinking isn’t supported. Do not send it.
Multi-turn reasoning and clear_thinking
Chat Completions is stateless. Resend conversation history with each request, including any assistant reasoning and tool calls required by the model. For tool calls, include the full assistant response followed by a matching result for each tool-call ID. See Tool Calling.
For Qwen 3.8 27B and GLM 5.3 Flash, clear_thinking controls whether the model receives historical assistant reasoning. It doesn’t remove the assistant’s final-answer content.
This example makes a Qwen request, includes its reasoning in the conversation history, and removes that reasoning from the follow-up prompt with
clear_thinking=True. To preserve the reasoning instead, omit clear_thinking or set it to False.
OpenAI client parameters
reasoning_effort is a standard OpenAI client parameter. Pass Cerebras-specific parameters such as reasoning_format and clear_thinking through extra_body when you use an OpenAI client. See OpenAI Compatibility.
