Nebius released SWE-rebench V2, a dataset of 32,000+ real software engineering tasks across 20 programming languages. It replaces the Python-only SWE-bench as the standard for training and evaluating AI coding agents.