Ant Group releases six Ling-3.0 base checkpoints, including WSM-merged stages

Article automatically generated from technical news.

Ant Group's Ling team has released six public base checkpoints: pretrained, mid-trained, and WSM-merged stages for each of Ling-3.0-tiny and Ling-3.0-flash. That is two sizes × three training stages. The repositories are ungated and declare MIT. None of the six checkpoints has been post-trained. The WSM-merged stage is the technical distinction in this Ling-3.0 base model release. In the WSM paper, Warmup-Stable-and-Merge keeps the learning rate stable after warmup, saves chec