Nemotron 3.5 Super VL lets you turn reasoning off, run it at medium effort, or run it fully on, and the reasoning tokens come out of the same max_tokens budget as the answer. Choose the level for each task and size the cap so the answer is never cut off.
NVIDIA's Nemotron 3.5 Super VL model card describes three reasoning levels: off, on with medium effort, and on. If you serve this model for image or video work and leave reasoning to chance, you will pay for tokens you do not need on easy jobs and risk truncated answers on hard ones. The fix is to pick the level per task and size the output budget to match.
The switch is a chat template argument. The card says reasoning is controlled through enable_thinking, and that setting enable_thinking to False disables reasoning. Medium effort is set with enable_thinking True plus medium_effort True. In the OpenAI compatible client that means passing extra_body with chat_template_kwargs holding those two keys. The card says medium effort will reduce the number of tokens generated compared to on, but may impact accuracy, so treat it as a cost dial you test, not a free upgrade.
The budget is the part people miss. The card says reasoning tokens and the final answer share the max_tokens budget. Its own code examples use a max_tokens of 16384 and raise an error when finish_reason equals length, which is NVIDIA telling you that a cut off reply is a failure to handle rather than a normal result. If you run reasoning on with a small cap, the model can spend the whole allowance thinking and never reach the answer.
Sampling needs one deliberate check. The card's summary table lists temperature 1.0 and top_p 0.95 for text and temperature 0.6 and top_p 0.95 for vision, while its code examples use temperature 1.0 and top_p 0.95 for image and video inputs as well. NVIDIA does not explain the difference, so do not guess. Run your own sample of vision requests at both settings and keep whichever gives more stable results.
Same week context: NVIDIA announced Nemotron for Science on November 16, 2026, built on this model, which makes these controls worth setting before you point it at lab images or document scans. The card lists the license as OpenMDW-1.1.
Try it today
- Pass extra_body with chat_template_kwargs and set enable_thinking explicitly on every request, instead of relying on a default.
- Use enable_thinking False for simple extraction or labeling, medium_effort True with enable_thinking True for routine questions, and full on only for hard problems.
- Set max_tokens high enough to cover reasoning plus the final answer, and treat finish_reason equal to length as an error to retry or escalate.
- For vision requests, test temperature 0.6 against temperature 1.0 with top_p 0.95 on your own samples before standardizing.