Empero AI’s Qwythos-9B-v2 Kills Looping Bug, Keeps 1M-Token Context

Empero AI released Qwythos-9B-v2, fixing a looping issue that affected 6.7% of greedy generations, now reduced to 0%. The update uses Final-Token Preference Optimization (FTPO) with 2,000 preference tuples, preserving the model’s 1M-token context and reasoning capabilities. Benchmarks show stable or improved scores on MMLU (83.8%), ARC-Challenge (96.4%), and GSM8K (93.6%), with a slight dip on HumanEval (77.4%).