HuggingFace released VKORE-46M-NPU on October 14, 2026, saying it demonstrates real backpropagation on an inference-only NPU. It is the company's first hardware-related update since the release of FINAL-Bench in 2025.
HuggingFace reported 0.99 correlation between NPU and GPU loss trajectories, measured on a controlled comparison protocol. That compares with no prior benchmark scores provided by the source.
VKORE-46M-NPU is built on a decoder-only transformer architecture with RoPE positional encoding and targets causal language modeling. Availability begins with the model release on the Hub, initially for researchers and developers.
"Backpropagation on an inference-only NPU is a real wall," said VIDRAFT_LAB SeaWolf-AI. The team emphasized the need for a reproducible validation protocol to confirm the training path.
The announcement follows the release of FINAL-Bench, which aims to provide a benchmark for NPU training. HuggingFace framed the significance as a step toward redefining the role of inference accelerators in training workflows.
HuggingFace did not say the method is open-source, and the core training technique remains patent-pending. Scaling to larger models and real-world data is explicitly the next step, not something claimed here.
Source: huggingface
