VKORE-46M-NPU

The source does not provide quantitative benchmark comparisons for VKORE-46M-NPU, including perplexity scores, downstream task accuracy, and training throughput in tokens per second.

VIDRAFT did not announce a public access method for VKORE-46M-NPU, with no Hugging Face repository, GitHub link, OpenAI-compatible API endpoint, or model weight download information.

VIDRAFT pre-trained a 46-million-parameter language model, VKORE-46M-NPU, from scratch on a class of NPU hardware normally used only for inference, with the entire training process not relying on NVIDIA GPUs or the CUDA software stack.

VKORE-46M-NPU is positioned as a proof of concept for hardware-agnostic AI development.

VIDRAFT's VKORE-46M-NPU pre-training reached convergence, that is, the model successfully completed training and did not abort due to hardware failure or numerical instability.

The publicly released VKORE-46M-NPU is a 46M-parameter, character-level model trained on the TinyStories dataset, and its purpose is to verify the feasibility of the training path rather than model quality.

Between NPU training and GPU training, the loss trajectory correlation coefficient reaches 0.99, which measures the similarity of training dynamics and does not mean throughput parity or that speed can catch up with GPUs.

VIDRAFT's release proves that backpropagation on an inference NPU is numerically credible, but the model has only 46M parameters, the training code is not open-sourced, and speed metrics were not disclosed; the engineering obstacles between 46M and larger-scale models currently cannot be judged.

The VKORE-46M-NPU model was trained on public datasets and can generate short English stories.

The core training technology of VKORE-46M-NPU is proprietary, is in the process of patent application, and details will be published in a subsequent paper; the core training code has not been open-sourced for the time being for intellectual property protection.

Under exactly the same initialization, data and learning rate schedule, the research team trained the same model from scratch on a 32-chip NPU and on a GPU respectively, and the Pearson correlation coefficient between the two loss curves reached 0.99.

This training ran on a first-generation inference-specific NPU released in 2025, using 32 chips, and is not the latest hardware.

VKORE-46M-NPU has a parameter scale of 46M, about one-third of GPT-2 Small (about 124 million parameters), and its significance lies in validating that the NPU training path is feasible, rather than in the quality of the model itself.

VIDRAFT released VKORE, an AI acceleration software layer for NPUs, which through software means gives NPUs originally capable only of inference the ability to train, without replacing hardware.

The model weights, model card, and comparative analysis materials have been made public on Hugging Face and can be obtained via huggingface-cli download vidraft/VKORE-46M-NPU; the core training code has not been made public.

The model weights, model card, and comparative analysis materials on NPU versus GPU training have been published on Hugging Face.

VIDRAFT released a small language model named VKORE-46M-NPU with 46 million parameters, pretrained entirely from random weights, and the training process did not use any GPU.

According to VIDRAFT, scaling to larger models and real data is the next step. But a single source points out that the core training technology has not yet been made public, which means independent reproduction cannot currently be completed.

The VKORE-46M-NPU model has been released on Hugging Face: FINAL-Bench/VKORE-46M-NPU, including weights, a model card and comparison data.

VIDRAFT completed full backpropagation on an inference-specific NPU, pretrained a language model from scratch, and released it as VKORE-46M-NPU.