Qwythos-9B-v2, a 9-billion-parameter text generation model by empero-ai built on the Qwen3.5-9B architecture, fixes repetitive output problems in the base Qwythos while preserving reasoning capabilities. Its key innovation is FTPO (Final-Token Preference Optimization), which eliminated the looping/repetition problem from v1, dropping from 6.7% to 0% looping under greedy decoding. The model ships with a native multi-token-prediction head and a 1,048,576-token context via YaRN rope-scaling. It achieves 83.8% MMLU with chain-of-thought, 93.6% GSM8K, and 77.4% pass@1 on HumanEval. The model maintains a 0% refusal rate on legitimate technical topics and allows greedy decoding or low-temperature sampling without relying on repetition_penalty.