Vidraft Pretrains Small Language Model on an Inference-Only NPU, No Nvidia GPUs

Vidraft said on October 15 that it pretrained a small language model using only an inference-only NPU, with no Nvidia GPUs. The company published the resulting model, VKORE-46M-NPU, on Hugging Face, according to aitimes.com.

The company said it found no prior case of the same training method in public literature and described the result as effectively a world first. Training ran on a first-generation inference-only NPU that Vidraft released in 2025.

Under the same conditions, a high-performance GPU run converged to equivalent loss curves and throughput, Vidraft said, with a loss trajectory correlation coefficient of 0.99. The released model was trained on public datasets and generates short English stories.

Vidraft published the model weights, a model card and a GPU training comparison analysis, but withheld the core training code to protect intellectual property. The core training technique has been filed for a patent and will be detailed in a future paper, the company said.

Vidraft added that it applied general public models to the NPU to verify inference acceleration, and confirmed that quantization improved processing speed. It is folding both the inference acceleration and the training technology into VKORE, an NPU-only software layer.