Some large models include their reasoning process in the output. This feature limits the model's thinking depth: when the reasoning exceeds the [object Object], the system truncates the chain of thought using a prompt, encouraging the model to stop reasoning early. This is useful for scenarios requiring a flexible trade-off between response speed and answer quality.
The Atlas 800I A2 inference server, Atlas 800I A3 SuperPoD server, and Atlas 300I Duo inference card support this feature.
Currently, only the Qwen3-32B, Qwen3-235B-A22B, and Qwen3-30B-A3B models support this feature.
To enable the thinking budget, include the following field in your request:
[object Object].
The value must be in the range[object Object].Currently, only the OpenAI inference interfaces are supported.
This feature currently does not support being enabled simultaneously with postprocessing parameters related to multi-sequence inference, such as
[object Object].
To enable the Thinking Budget feature, the parameters that need to be configured are shown in .
Table 1 Thinking Budget parameter: [object Object] in [object Object] [object Object][object Object]
Open the
[object Object]file of the server.- Installation using the
[object Object]package:
[object Object]- Installation using the
[object Object]package:
[object Object]- Installation using the
Configure the serving parameters. Add the
[object Object]field to the Server's[object Object]file according to . For details on serving parameters, see . A parameter configuration example is shown below.Qwen3-32B is used as an example:
[object Object][object Object]
Start the service. For PD colocation scenarios, refer to "Quick Start" > "" in MindIE Motor Developer Guide. For PD disaggregation scenarios, refer to "Cluster Service Deployment" > "" in MindIE Motor Developer Guide.
Send a request. For parameter descriptions, see the "" section in MindIE LLM Developer Guide