Bdraft (비드래프트) releases 'VKORE-46M-NPU' on Hugging Face, pretrained from scratch using only inference NPUs (report)

Verified (fact)

According to ZDNet Korea, Bdraft released 'VKORE-46M-NPU' on Hugging Face, a small language model trained without GPUs using only first-generation inference-only NPUs launched in 2025. The company said it performed backpropagation on the NPUs with its own technology and trained the model from scratch without pretrained weights. The model was trained on public datasets and can generate short English stories. The weights, model card and comparative analysis were released, but the core training code remains private.

Needs fact-check (claim)

The company said the correlation coefficient between its loss curve and that of a GPU training run under the same conditions was 0.99. This only means the training processes were similar, not that training speed was the same.

Why it matters (interpretation)

This is the first public result showing that the VKORE stack covered in earlier briefings has actually performed training. However, the model is only 46M parameters and the training code is private, making external reproduction difficult.

Sources

• Press 비드래프트, GPU 없이 NPU만으로 AI 사전학습 성공 ZDNet Korea

Story thread

1. 2026-10-14 Bdraft (비드래프트) unveils 'VKORE,' an acceleration software layer for running both inference and training on NPUs (report)
2. 2026-10-14 Bdraft (비드래프트) releases 'VKORE-46M-NPU' on Hugging Face, pretrained from scratch using only inference NPUs (report)

More about Bdraft

• 2026-10-14 P2 Infrastructure Announced KoreaEfficiency & infra Bdraft
  Bdraft (비드래프트) unveils 'VKORE,' an acceleration software layer for running both inference and training on NPUs (report)

  Bdraft unveiled 'VKORE,' an acceleration software layer that improves NPU inference performance and also enables training. It works like...