Phi small language models from Microsoft Research — efficient reasoning at sizes that fit on-device or on a single GPU.