I am an individual user testing the free OpenRouter endpoint of Nemotron 3 Ultra 550B through Chub AI, and I am reporting a reproducible generation/formatting problem.
I’m experiencing a potentially related issue with Nemotron 3 Ultra 550B on the OpenRouter free endpoint. My issue is different from the multilingual/token-soup corruption described here: after roughly 10–20 messages, the model can progressively develop persistent * formatting and continues doing so despite explicit in-chat/OOC instructions, assistant prefill attempts, and manually correcting previous responses. I’ve opened a separate thread with the full reproduction details.