“The NVIDIA Llama-Nematron reasoning paper is considered the most detailed public resource on training a hybrid reasoning model.”