Microsoft’s small language model outperforms larger models on standardized math tests