According to CoinMeta, Ant Birding has officially open-sourced the weights of the ling-3.1-flash-1019 model, and also released the ling-3.1-flash-base model. Both models have been launched on Hugging Face and Modelscope, and are licensed under the MIT open-source license. The new versions continue to adopt a hybrid expert architecture with a total of 560B parameters and 25B activation parameters, mainly enhancing inference through post-training and agent capabilities. The artificial intelligence index has increased from 41 points in the old version to 44 points. In evaluations of complex office tasks, AA-briefcase has improved from around 1400 to 1524 ELO, in software operation evaluations, Automationbench-AA has increased from 62% to 68%, and in physical reasoning tests, Critpt has increased from 18% to 26%. The simultaneously opened base models are released with BF16 accuracy, allowing developers to continue with pre-training, industry-specific fine-tuning, or reinforcement learning using them.