Alibaba open-sources AI software stack SAIL: 260+ frameworks ready out of the box

At WAIC 2026, Alibaba's T-Head announced the open-sourcing of its proprietary AI software stack SAIL (Seed of AI Library). SAIL is the underlying software for the Zhenwu series AI chips, covering the full chain from OS layer, SDK layer to interface layer, with independent intellectual property rights and full compatibility with mainstream AI ecosystems. The open-source release includes source code, drivers, tools, and documentation, available on the T-Head developer community. As of April 2026, Zhenwu AI chips have shipped over 560,000 units, serving more than 400 customers across 20+ industries, and have been deployed in Alibaba Cloud and various enterprise production environments, surviving massive traffic peaks like Double 11. The stack comprises five layers: Driver & Runtime (PPU Runtime and PPU Driver), Programming Language (C/C++ and Python interfaces), Compiler (Compiler and Debugger), High-Performance Libraries (math libraries like acBLAS, acFFT, acDNN; codec libraries HGENC, HGDEC; communication libraries PCCL and sailSHMEM; extension library acext and ML library HGML), and Development Tools (Asight Compute for operator-level profiling, Asight Systems for system-level profiling, PPU-SMI for device monitoring, PPU-DCGM for cluster resource scheduling). SAIL supports over 260 mainstream training and inference frameworks including PyTorch, TensorFlow, vLLM, and SGLang, with an average adaptation time of less than 7 days for new models. T-Head aims to build an ecosystem by lowering the entry barrier and inviting global developers, research institutions, and industry partners to co-develop tools, enrich the operator ecosystem, and optimize performance.