Four AI agents reportedly post-trained a Qwen base model to instruct-level scores in 104 hours

Brian Roemmele describes Athanor research in which four Claude Opus 5.5 agents post-trained Qwen3-4B-Base in 104 hours. He reports final scores of 67.3 on a six-task development suite and 48.2 on a held-out suite, versus 66.5 and 47.9 for Qwen3-4B-Instruct-2507. The agents called no external teacher models, though some public training data had been distilled from stronger models. The result comes from one campaign and trailed Qwen's reasoning releases.