Ling-3.0 opens six base checkpoints across three training stages

Ant Group's new release makes six base checkpoints available: pretrained, mid-trained, and WSM-merged checkpoints for both the tiny and flash sizes. That is two sizes × three training stages. Every repository is public and ungated with an MIT declaration, and none of the six has been post-trained.

The Ling-3.0 base model release exposes three points in the training progression for each of two model sizes. Researchers can inspect each released stage, but these are base checkpoints for continued pretraining, fine-tuning, and research—not finished chat or instruct models.

One important evidence boundary: the team says the same recipe was validated on tiny and then scaled to flash, but that is an official statement rather than an independent reproduction. The WSM paper's reported experiments use Ling-mini, not the six Ling-3.0 checkpoints released here.

The official release announcement contains the full family, and the main tiny and flash model pages are the WSM-merged endpoints..