phi-1-5

Microsoft's 1.3B Phi-1.5 model — trained on synthetic 'textbook quality' data, outperforms much larger models on reasoning tasks.