Fully open releases and reproducible AI research raise the transparency bar

NVIDIA’s open agentic-coding midtraining recipe reports SWE-bench gains for Nemotron 3.5 Lightning, with coverage citing results from 55.8 to 62.8 and elsewhere 64.5; the figures and evaluation details need reconciliation and independent replication. NeMo-Run documentation now describes execution across local, Docker, Slurm, DGX Cloud, and Lepton environments. Community catalogs and model-specific fine-tuning recipes provide practical leads, but claims and workflows still need artifact-level validation.