VIDRAFT trains language model from scratch on inference-only NPU chip

VIDRAFT has successfully pretrained a small language model entirely on an inference-only NPU chip, challenging the industry assumption that such chips cannot perform backpropagation. The resulting VKORE-46M-NPU model was trained under identical conditions on both the NPU and a GPU, with the two loss trajectories showing a 0.99 correlation. This demonstrates that the NPU performed numerically equivalent training work despite lacking the hardware paths typically considered essential. The model and its weights have been published on Hugging Face, while the proprietary core training technique remains private for IP protection.