1.5B Open-Source VLA: Efficiency Breakthrough or Data Hallucination?

I noticed an interesting detail: the MiniCPM-Robot series models released by ModelBest have a parameter scale of only 1.5B, yet they claim to "break into the first tier of embodied intelligence." In quantitative finance, we're used to evaluating strategies with metrics like Sharpe ratio and maximum drawdown. When facing a model, my first reaction is: What is its training data volume? What are the evaluation benchmarks? How is its out-of-sample generalization ability? If this information is missing, any "first tier" label is just marketing alpha, not actual technical alpha.