TorchAO is the official PyTorch library for model optimization through quantization and sparsity. It provides techniques including float8 training for 1.5x pre-training speedup, quantization-aware training (QAT) to recover accuracy lost in post-training quantization, and int4 weight-only quantization achieving 1.89x inference speedup with 58% memory reduction. The library supports transformer models like Llama and Gemma and includes optimizations for mixed-expert (MoE) architectures using MXFP8 precision.