Huawei Ascend Achieves Day-Zero Adaptation for Ant Group's Bailian Ling-3.0-Flash Model

On August 6, Ant Group's Bailian open-sourced Ling-3.0-flash, a new generation of native hybrid inference model. On the same day, Huawei announced that Ascend had completed Day-0 (0Day) adaptation for this model and introduced, for the first time, the CANNPyPTO operator programming framework to shorten the delivery cycle of complex fused operators and enhance the efficiency of operator integration and inference framework performance tuning. This adaptation fully supports Ascend A2 and A3 series products and provides an official standardized deployment solution, enabling developers to quickly launch inference services using official container images. Built upon PTOISA, CANNPyPTO allows developers to define computation logic via tensor-level APIs, with the framework automatically handling sharding scheduling, data movement, and code generation. Additionally, CANNBotPyPTOAgent establishes a multi-agent framework that divides operator development into seven stages, each executed collaboratively by specialized sub-agents. Through a unified orchestrator and memory mechanism, these agents exchange information and pass context seamlessly—transforming natural language operator requirements into board-ready operators that meet both accuracy and performance targets, all without any manual intervention.