Nemotron 3 Super hits reasoning loop

This smells like prompt plumbing gone wrong, not the base weights suddenly losing it. The fact that the same model is fine in some frontends but loops in others is the biggest clue.

• –The poster says context was not overflowing, so this does not look like a simple window-limit failure.
• –Other commenters reproduce similar behavior in different stacks, while one notes it works in LM Studio, which strongly suggests backend-specific formatting differences.
• –Nemotron 3 Super is tuned for agentic reasoning, so if a runtime leaks reasoning traces back into the next turn, the model can self-amplify into a loop.
• –A bad quant or GGUF could worsen the problem, but cross-backend variation makes chat-template and stop-token handling the first thing to debug.
• –For local deployments, this is a reminder that reasoning models can be unusually sensitive to how “thinking” text is captured and reinserted.