VIDRAFT trains 46M-parameter model entirely on inference-only NPUs without CUDA

Korean startup VIDRAFT has fully pre-trained VKORE-46M-NPU, a 46-million-parameter language model, using only inference-oriented NPU hardware, bypassing Nvidia GPUs and the CUDA stack. The company frames it as a proof-of-concept for hardware-agnostic AI development.

• VKORE-46M-NPU has 46 million parameters
• Training ran on inference-only NPUs, no Nvidia GPUs or CUDA
• It was full pre-training, not fine-tuning or distillation
• No benchmarks or model weights have been released