Choose model files or benchmarks, then stream the run output as it arrives.