# Qwen

ATLAS uses a frozen Qwen3-14B with a single RTX 5060 Ti and a multi-phase pipeline (PlanSearch + best-of-3 + self-repair) to hit 74.6% on LiveCodeBench — passing Sonnet 4.5's 71.4%. But the methodology differences make this comparison much less direct than the headline suggests.