NVIDIA NIM Thinking Budget Control: nvext.max_thinking_tokens Caps Reasoning Tokens Per Request Without Model Changes

NVIDIA NIM Docs·medium signal

NVIDIA NIM now exposes a `nvext.max_thinking_tokens` field on chat completion requests that hard-caps how many thinking tokens a reasoning model generates before it must produce its final answer. Set `NIM_ENABLE_BUDGET_CONTROL=1` plus model-specific start/stop thinking tag env vars; Nemotron-Nano-9B-v2 ships with it on by default. This decouples cost and latency control from model selection — you can run the same model at "fast and cheap" for simple tasks and "slow and thorough" for complex ones by varying the budget per request, without switching endpoints or models.