VIDRAFT's 325M-Parameter Model Pretrained on an Inference NPU Matches GPU Loss

VIDRAFT has scaled its earlier 46M-parameter model to VKORE-325M-NPU-Base, a 325.6M-parameter language model trained from random initialization on Rebellions' ATOM-Max inference NPU. A twin model trained on an NVIDIA B200 under identical settings served as the reference. On 1,024 windows of unseen FineWeb-Edu text, the NPU model scored a loss of 0.7993 against 0.7994 for the GPU model. The run covered 5 billion tokens over 76,292 steps, about four days, and finished without divergence, with the training curves overlapping throughout. The researchers credit VIDRAFT's own software stack for making training on the chip possible, and both models' weights and the evaluation script have been made public.