Modes
- ON: keep chain-of-thought for multi-step reasoning, tools, math.
- OFF: concise replies, better for chat and fan-out agents.
Reasoning control
Reasoning ON/OFF and Thinking Budgets in Nemotron 3
Control chain-of-thought depth with Reasoning ON/OFF and thinking-token budgets to balance accuracy, privacy, and cost.
nemotron reasoningthinking budgetreasoning on offchain of thought controlnemotron budgets
Budgets
- Declare max thinking tokens in prompt/request; budget caps depth.
- Use with ON to balance quality and cost.
Best practices
- Default OFF for frequent calls; toggle ON per task with budgets.
- Safety: OFF reduces CoT leakage in sensitive contexts.
- Monitor thinking tokens and refine prompts/limits over time.
FAQ
How to set budgets via API?
Include a max thinking-token limit in the prompt or request payload.
Does OFF reduce quality?
Minimal impact on simple Q&A; for complex reasoning use ON with budgets.
Can I switch per request?
Yes. Toggle ON/OFF and budgets per call based on scenario.