“Gemma 4's performance on human-rated tasks is competitive with larger, frontier open models.”