How do I turn off thinking in Ollama?

Get a direct answer from a supported model

Use --think=false when starting an Ollama model, or enter /set nothink inside its interactive terminal chat. This requests the model's non-thinking behaviour where supported. --hidethinking only hides the displayed trace: it does not make the same request. Ollama documents these separate controls in its thinking announcement.

For quick rewriting, classification or a short everyday question, a direct answer can be more convenient than a long reasoning trace. Keep thinking available for tasks where you want to compare the result. There is no need to rebuild a model or buy another app just to try the setting.

Start with Ollama already installed and a model already available on your computer. These instructions were source-checked on 12 October 2026; no inference or speed benchmark was executed for this answer.

1. Check the exact model and your version

Open a terminal and run:

ollama --version
ollama ls
ollama run --help

Copy the complete model tag from the list, including its size or suffix. Use that tag wherever you see YOUR_EXISTING_MODEL_TAG below. Check your installed CLI advertises --think and --hidethinking. The CLI reference covers listing and running models; the current command source distinguishes the two flags.

Do not assume every tag in a family has identical controls. A thinking-only variant and an instruct variant can behave differently. If the command starts an unexpected download, cancel and check the name. For a local run, confirm you selected local weights rather than a cloud entry using our local-or-cloud check.

2. Turn thinking off in the terminal

For one question, substitute your existing model tag:

ollama run YOUR_EXISTING_MODEL_TAG --think=false "Rewrite: We will get back to you soon."

For an ongoing chat, start ollama run YOUR_EXISTING_MODEL_TAG, then type these lines at the >>> prompt, one at a time:

/set nothink
Rewrite: We will get back to you soon.

To request thinking again in that chat, enter /set think when the model supports it. These commands change the terminal session's requests; they are not a computer-wide switch for every frontend. The interactive CLI source handles them separately from ordinary prompts.

If you want the model to keep its thinking behaviour but show only its answer, use --hidethinking instead. You can copy the answer from the terminal, or use the explicit file export below.

3. Check which controls your model accepts

Current Ollama documentation exposes supported controls through POST /api/show. With your local Ollama server available at its usual port, open PowerShell and paste:

$thinkModel = (Read-Host 'Exact existing model tag').Trim()
$thinkApi = 'http://127.0.0.1:11434'
$thinkInfo = Invoke-RestMethod -UseBasicParsing -Method Post `
    -Uri "$thinkApi/api/show" -ContentType 'application/json' `
    -Body (@{model = $thinkModel} | ConvertTo-Json)
$thinkInfo.thinking | ConvertTo-Json -Depth 4

Read values and default, as described in model details:

• false and true: the model advertises on/off controls.
• Named levels, such as low, medium and high: use a listed name exactly. GPT-OSS is the documented example; low is a lower level, not an off switch.
• Only false: the model does not support thinking.
• No thinking object: metadata is missing. That alone does not establish whether it produces a reasoning trace.

The thinking documentation says unsupported named levels can fall back to the model default. Avoid guessing none, minimal or another application's setting name for the native API. If your version lacks this metadata, check its help and the exact model's current documentation before choosing a mode. Update Ollama separately if necessary.

4. Save only the answer with PowerShell

If your model supports false, continue in the same PowerShell window. This sends one simple request, then saves the final text as ollama-direct-answer.txt in your current folder. Use a fresh folder if that filename already exists.

$ErrorActionPreference = 'Stop'
$thinkOutput = Join-Path (Get-Location).Path 'ollama-direct-answer.txt'
if (Test-Path -LiteralPath $thinkOutput) { throw 'Output already exists.' }
$thinkBody = @{
    model = $thinkModel
    think = $false
    stream = $false
    messages = @(@{
        role = 'user'
        content = 'Rewrite: We will get back to you soon.'
    })
} | ConvertTo-Json -Depth 5
$thinkResult = Invoke-RestMethod -UseBasicParsing -Method Post `
    -Uri "$thinkApi/api/chat" -ContentType 'application/json' `
    -Body $thinkBody -TimeoutSec 300
if ($thinkResult.done -ne $true -or
    [string]::IsNullOrWhiteSpace($thinkResult.message.content)) {
    throw 'No complete text answer returned. Inspect the response.'
}
$thinkBytes = [System.Text.Encoding]::UTF8.GetBytes($thinkResult.message.content)
$thinkFile = [System.IO.File]::Open($thinkOutput, [System.IO.FileMode]::CreateNew)
try { $thinkFile.Write($thinkBytes, 0, $thinkBytes.Length) }
finally { $thinkFile.Dispose() }
Get-Content -LiteralPath $thinkOutput -Encoding UTF8

think belongs beside model and messages, not inside options. stream = false requests one JSON response, and the chat API puts the answer in message.content. The script deliberately exports that field rather than message.thinking. Microsoft documents REST requests and CreateNew, which refuses to replace an existing file.

When the setting seems to do nothing

It works in the terminal but not your chat app. Check that app's model/request settings. An Ollama CLI session does not configure another frontend. A collapsed thinking panel may simply hide text. The UI feature discussion shows why an identical toggle cannot be assumed across versions and interfaces.

You are using GPT-OSS or a named-level model. Inspect its supported controls first. Lower effort, hidden output and a genuinely non-thinking model are different choices. Ollama's OpenAI compatibility interface has its own reasoning_effort conversions; do not copy those fields into the native /api/chat request. Missing visible reasoning does not establish that all reasoning computation stopped.

A Modelfile or prompt trick fails. PARAMETER thinking false is not a documented Modelfile parameter in the reference. Qwen's original Qwen3 instructions describe /no_think as a model-specific soft prompt switch; it is not a universal Ollama command. Prefer the native controls above.

The answer still takes too long. Loading, model size and long conversation history remain separate costs. Start a fresh short chat and compare the same prompt with each supported mode, recording whether each answer actually completes and is useful. See why local AI runs slowly for the wider checks.

Local or cloud?

An existing local model lets you try these controls without a hosted subscription, using your own hardware. Ollama's cloud route can suit a computer that cannot comfortably run the model, with account limits and optional paid plans. It does not create an off switch the selected model lacks. Solve the mode-setting issue first; consider hosting only if hardware remains the obstacle.