Error: 400 Bad Request: "qwen2.5:0.5b" does not support thinking — you asked for thinking mode, with ollama run --think or "think":true in the API, on a model whose capabilities do not include thinking. Drop the flag for that model, or switch to a model whose capabilities include thinking, such as Qwen 3.

Run ollama show on the model and read its Capabilities list. If thinking is not in it, no flag turned it on in the test; the server refused the request with HTTP 400.

The error

PS> ollama run qwen2.5:0.5b --think "What is 17 times 3?"
Error: 400 Bad Request: "qwen2.5:0.5b" does not support thinking
exit code: 1

The API gives the same answer to "think":true:

PS> curl.exe -s -i http://127.0.0.1:15190/api/generate -d '@think.json'   (think.json: {"model":"qwen2.5:0.5b","prompt":"What is 17 times 3?","think":true,"stream":false})
HTTP/1.1 400 Bad Request
Content-Type: application/json; charset=utf-8
Date: Fri, 06 Nov 2026 07:21:08 GMT
Content-Length: 54

{"error":"\"qwen2.5:0.5b\" does not support thinking"}

Why it happens

Thinking models such as Qwen 3 write a reasoning section before the answer, and Ollama separates it into its own thinking field. ollama show prints a Capabilities list for every model, and in the test the error matched that list exactly: the model that lists thinking accepted the flag, the model that does not was refused. qwen2.5:0.5b lists only completion and tools: 

>   Capabilities
      completion
      tools

Asking for a level does not help: --think=high gave the identical error. The flag's own help text, from ollama run --help, limits it to supported models: 

      --think string[="true"]   Enable thinking mode: true/false or high/medium/low for supported models

The fix

For a model without thinking, leave the flag off and you get a plain answer:

PS> ollama run qwen2.5:0.5b "What is 17 times 3? Answer with the number only."
51

exit code: 0

On qwen2.5:0.5b, --think=false was accepted (it answered 17 * 3 = 51). If you want the reasoning, use a model that has the capability. qwen3:0.6b lists it, with the levels it accepts: 

>   Capabilities
      completion
      tools
      thinking
          levels     false, true
PS> curl.exe -s http://127.0.0.1:15190/api/generate -d '@think3.json'   (think3.json: {"model":"qwen3:0.6b","prompt":"What is 17 times 3?","think":true,"stream":false,"options":{"num_predict":600}})
PS> $j.thinking.Length; $j.response; $j.done_reason
1024
To calculate $17 \times 3$, we can break the multiplication into simpler steps:
...
stop

The answer came back with 1,024 characters of reasoning in thinking, a final answer of 51 in response and done_reason stop. A word of warning from the same test: the same tiny model run with ollama run --think and no token limit kept reasoning in a loop and never reached an answer, so cap num_predict when you try thinking on very small models.

How it was reproduced

Ollama 0.40.14 on Windows 11, against a separate test server started with ollama serve on 127.0.0.1:15190 with its own models folder. qwen2.5:0.5b and qwen3:0.6b were pulled into it. The CLI was run with --think, --think=high and --think=false, and the API with JSON bodies sent by curl.exe; the qwen3 answer was read with ConvertFrom-Json in PowerShell.

Frequently asked

Why does Ollama say a model does not support thinking?
You passed --think or think true for a model whose capabilities, as ollama show prints them, do not include thinking. The server answered 400. Use a thinking model or drop the flag.
How do I know if an Ollama model supports thinking?
Run ollama show with the model name. A thinking model lists thinking under Capabilities, with the levels it accepts; qwen3:0.6b showed false and true, qwen2.5:0.5b showed no thinking at all.
Does --think=false work on a model without thinking?
It did in the test with Ollama 0.40.14: qwen2.5:0.5b, which has no thinking capability, accepted --think=false and answered normally, while --think and --think=high were refused with the 400 error.

More decoded errors in the Fixes category. A similar error refuses tool calls on models without tools, covered in 'does not support tools' from .NET.