Those numbers are for cost only, not performance. It’s not clear they actually trained a 70B vs. just using randomly initialized parameters.