Reasoning
What is it?
Reasoning models are trained to produce an explicit thinking phase before the final answer — either always ("thinking" models) or when prompted into a reasoning mode.
When should I use one?
- Math, logic puzzles, multi-step analysis, careful planning.
- Tradeoff: thinking tokens take time and context. For simple chat, a regular chat model is faster.
Prerequisites
- A reasoning-tagged model — filter the catalog by Reasoning.
- Sufficient context length for the thinking phase plus the answer. See Context and memory.
Minimal working example
bash
curl http://127.0.0.1:11435/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"<publisher>/<model>","messages":[{"role":"user","content":"Solve: if 3x+7=22, what is x?"}]}'The response contains the model's reasoning followed by the answer. How the thinking phase is delimited depends on the model's chat template.
Configuration/options
- Some models toggle reasoning with a chat-template flag or a system prompt; check the model card.
- Thinking output may appear in-band with the answer depending on template support; parsing that into separate fields is planned. Planned
Limitations
- Model-dependent behavior: templates vary between model families; some emit
<think>blocks, others plain text. - Thinking tokens consume context — very long reasoning can truncate.
- Do not confuse "reasoning badge" with guaranteed correctness; verify critical outputs.
Troubleshooting
Thinking never ends, or answer mixed into reasoning: template mismatch — see Model problems.