View reasoning information

How to return and view reasoning in your Serverless Inference responses

Reasoning models, such as Google’s Gemma 4, return information about their reasoning steps alongside the final answer. This page explains how to identify reasoning-capable models on Serverless Inference, where to find reasoning output in a response, and how to turn reasoning on or off for models that support toggling it. Use this guide to inspect a model’s intermediate reasoning or to control whether reasoning appears in a response.

To determine whether a model supports reasoning, check the following Supported models table or the Supported Features section of its catalog page in the UI.

Reasoning information appears in the reasoning field of responses. The value of this field is null in the responses of non-reasoning models.

Supported models with reasoning

The following table lists the models on Serverless Inference that return reasoning output. Each supported model either always includes reasoning, or disables or enables reasoning by default:

Model ID (for API usage)	Reasoning support
deepseek-ai/DeepSeek-V4-Flash	Disabled by default
deepseek-ai/DeepSeek-V4-Pro	Disabled by default
google/gemma-4-31B-it	Disabled by default
MiniMaxAI/MiniMax-M2.5	Always on
moonshotai/Kimi-K2.6	Always on
moonshotai/Kimi-K2.5	Always on
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8	Enabled by default
openai/gpt-oss-120b	Always on
openai/gpt-oss-20b	Always on
Qwen/Qwen3.6-35B-A3B	Enabled by default
Qwen/Qwen3.6-27B	Enabled by default
Qwen/Qwen3.5-35B-A3B	Enabled by default
Qwen/Qwen3.5-27B	Enabled by default
Qwen/Qwen3-235B-A22B-Thinking-2507	Always on
zai-org/GLM-5.1	Enabled by default

Models with Always on reasoning

If a model is listed as Always on in the preceding Supported models table, it always includes reasoning, and you can’t disable it.

Disable reasoning

If a model is listed as Enabled by default in the preceding Supported models table, you can disable reasoning to reduce token usage or simplify the response. To opt out of reasoning for a request, in chat_template_kwargs, set the enable_thinking flag to False (Python) or false (Bash). After the request completes, the response omits the reasoning content:

Python

import openai

client = openai.OpenAI(
    base_url='https://api.inference.wandb.ai/v1',
    api_key="[YOUR-API-KEY]",  # Create an API key at https://wandb.ai/settings
)

response = client.chat.completions.create(
    model="google/gemma-4-31B-it",
    messages=[
        {"role": "user", "content": "3.11 and 3.8, which is greater?"}
    ],
    extra_body={
        "chat_template_kwargs": {
            "enable_thinking": False
        }
    },
)

Bash

curl https://api.inference.wandb.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer [YOUR-API-KEY]" \
  -d '{
    "model": "google/gemma-4-31B-it",
    "messages": [
      { "role": "user", "content": "3.11 and 3.8, which is greater?" }
    ],
    "chat_template_kwargs": {"enable_thinking": false}
  }'

Enable reasoning

If a model is listed as Disabled by default in the preceding Supported models table, you can enable reasoning by setting the enable_thinking flag to True (Python) or true (Bash) in the preceding code snippet.