Gustavo
0bb30696de
Nemotron spends the token budget on reasoning_content before answering; a 2048 budget was exhausted by ~7761 reasoning chars, returning empty content with finish_reason=length. Topic selection and category choice go 2048 -> 8192, image plan 4000 -> 8192, article rewrite 8192 -> 16384. max_tokens is only a cap, so concise answers cost the same as before.