4 × NVIDIA H100 (94 GB), one node of the Capella HPC cluster
Thinking
Default: on, with the highest effort (max).
reasoning_effort: every value is accepted, but the model only knows low and high.
All other values (minimal, medium, xhigh, max) mean max.
low and high both think clearly less than the default.
Numeric effort: no.
Turn off: not possible. With reasoning_effort: "none", chat_template_kwargs: {"enable_thinking": false}
or chat_template_kwargs: {"thinking": false} the model still thinks, and the reasoning ends up in the answer
(content) instead of the separate reasoning field (vLLM issue #54744).