VIDRAFT's Darwin Family Framework Achieves 86.9% on GPQA Diamond Without Retraining

Korean AI startup VIDRAFT has released a preprint introducing Darwin Family, a training-free model merging framework that improves large language model reasoning by combining parameters from two existing models using evolutionary algorithms. The framework pairs a generalist 'Father' model with a reasoning-specialized 'Mother' model, both sharing the same pre-training base, and evolves optimal merge configurations without any gradient updates. A 14-dimensional genome vector controls layer-level mixing ratios, while a diagnostic scoring system called MRI identifies which layers contribute most to reasoning performance. The flagship output, Darwin-27B-Opus, scored 86.9% on the GPQA Diamond benchmark, ranking 6th globally among 1,252 evaluated models at the time of publication. The research was published in May 2026 as arXiv preprint 2605.14386.