Ant Group’s Bailing large-scale model has officially released its open-source iteration, Ling-3.1-flash-1019. This version builds upon the foundational architecture of Ling-3.1-flash, achieving notable enhancements across multiple evaluation metrics. A key highlight is its significantly improved capacity to manage intricate workflows and multi-step tasks within Personal Agent applications. In parallel, the open-source base model, Ling-3.1-flash-base, employs a hybrid linear Mixture of Experts (MoE) framework, boasting a total of approximately 560 billion parameters, with around 25 billion parameters activated per token. This design ensures exceptional performance across a range of benchmark tests. Developers can now access the model under the BF16 base weights and MIT license. This release also aligns with the technical specifications of Ant Group’s previously launched personal AI assistant, Lingguang 2.0, serving as its foundational model. Furthermore, Bailing has developed a visual-language model, Ling-3.1-flash-VL, based on this series, with plans for external release in the near future. Bailing remains committed to advancing model development tailored to Personal Agent requirements, incorporating real-world feedback to refine alignment training. The initiative has already yielded specialized open-source models for the finance and healthcare sectors, facilitating the practical deployment of AI capabilities in real-world scenarios.