3/ Each model was pretrained on roughly 20T tokens, according to IFM.

The lab says it is publishing final weights, checkpoints, logs, evaluations, code, data where licensing permits, and detailed construction recipes where redistribution is restricted.