On the same day NVIDIA CEO Jensen Huang posted his first X post endorsing open-source AI models alongside 25 industry signatories, Ant Group Bailing division released Ling-3.0-flash, a 124B total parameter execution model with only 5.1B activated parameters that achieves higher intelligence density than its own trillion-parameter flagship. The model scored first place in 15 of 34 evaluation dimensions and second in 19, tying DeepSeek V4 Flash for the top average score of 67.6, while surpassing its own Ring-2.6-1T trillion-parameter model by nearly 8 points. Ling-3.0-flash achieves an intelligence density score of 0.5 and efficiency ratio of 13.2, both highest among comparable models. In SWE-Bench Pro it scored 56.6% first place, in ArtifactsBench it achieved 77.0%, in MiniAppBench it reached 25.3%, and BFCL-v4 function calling accuracy reached 73.0%. The model natively supports 256K context windows, with MRCR_256K at 81.1%, Multi-IF at 87.7%, LIFEBench at 77.3%, and BrowseComp at 82.0%. Ant Bailing open-sourced the model with weights available after the August 3 free API trial period ends.