Bonsai 27B field guide
Ternary Bonsai 27B needs room for reasoning tokens
Why a short output limit can return reasoning without a final answer, with a traced 16 token and 256 token comparison.
Why did Ternary Bonsai 27B return no final answer?
The model can use the whole output budget for reasoning before it writes a final answer. In our July 15 trace, a 16 token limit produced 15 reasoning tokens and no final content. A 256 token limit used 191 reasoning tokens and then returned the correct answer, 42.
The traced check
We asked the same simple addition question twice through LM Studio's OpenAI compatible API. Temperature was zero. Only the output token limit changed.
The small limit ended inside the reasoning field. The larger limit completed the reasoning and wrote the requested number in the content field.
What applications should do
- Budget reasoning tokens separately from the visible answer when the runtime reports them.
- Treat an empty content field as an incomplete response, even when token usage is nonzero.
- Test the shortest output limit that works for each task type.
- Check whether the selected runtime offers a tested thinking control before relying on it.
Questions people ask
Did the model fail the math question?
No. It reached the correct answer when the output budget was large enough.
Is 256 tokens always enough?
No. Harder tasks may need more reasoning. Measure the tasks your application sends.
Primary sources
Vendor claims on this page are labeled as PrismML claims. Local results are added only after a saved trace completes.