Ant Bailing has open-sourced the enhanced version of Ling-3.1, alongside its 560B parameter pre-training base model.

Beating AI Insight News Brief: Ant Group’s Bailing (Ling) model has officially open-sourced the Ling-3.1-flash-1019 model weights, and also released the Ling-3.1-flash-base model. Both models are now available on Hugging Face and ModelScope, under the MIT open-source license. The 1019 version was first announced on October 22, with only its API accessible at that time. The new version retains the Mixture of Experts (MoE) architecture with 560B total parameters and 25B activated parameters, and mainly enhances reasoning and agent capabilities via post-training. The Artificial Analysis intelligence score rose from 41 points in the previous version to 44. The complex office task benchmark AA-Briefcase improved from ~1400 to 1524 Elo, the software operation benchmark AutomationBench-AA increased from 62% to 68%, and the physical reasoning test CritPt rose from 18% to 26%. The concurrently released Base model is a foundational model that has not undergone full instruction alignment and post-training, and is released with BF16 precision. Developers can use it for continued pre-training, industry-specific fine-tuning, or reinforcement learning, without starting from an already aligned chat model.