VIDRAFT

VKORE aims to rebuild a CUDA-like abstraction layer on NPUs, providing a unified training and inference interface upward and adapting to different vendors' NPU instruction sets downward, so that data centers can swap accelerators like swapping storage or network cards without rewriting model code, thereby breaking single-vendor lock-in.

VKORE is positioned as a hardware-agnostic software layer, aimed at data centers with mixed GPU and NPU deployments, and is not tied to a specific chip vendor.

FINAL-Bench has a technical partnership with the leaderboard maintainer QuantID, and its parent company VIDRAFT appears in the rankings, so this daily report is not an independent evaluation.

VIDRAFT's Darwin product line has multiple models at different scales and inference configurations near the top of the GPQA leaderboard, including 180B-RSI (first), 397B-ZTC (third), and 398B-JGOS (ninth, 90.9%), indicating a multi-version parallel iteration pipeline.

Under the AIhoster methodology, Darwin-180B-RSI's scores come from VIDRAFT's own model card report and have not been independently reproduced by a third party.

The publisher QuantID is a technology partner of VIDRAFT, and FINAL-Bench is an organization under VIDRAFT, so the leaderboard involves a conflict-of-interest disclosure, and the original text states that this is not an independent evaluation.

According to VIDRAFT, cumulative downloads of POCKET's various builds have exceeded two million.

FINAL Bench is a functional metacognition evaluation benchmark covering 15 domains and 100 tasks, aimed at assessing models' ability to identify their own errors and correct them.

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

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

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.

The experiment ran the same model, same initialization, same data and same training schedule on a 32-chip NPU cluster and a reference GPU, and the correlation between the NPU and GPU loss trajectories reached 0.99.

VIDRAFT additionally defined a compilation path for the backward graph at the software layer for the NPU, so that gradient computation can also be compiled into static programs the NPU can execute.

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.

The VIDRAFT team completed full pretraining of a language model from scratch on an inference-only NPU, without using any GPU.

The backpropagation implemented by VIDRAFT on the NPU is standard reverse-mode gradient computation, not a gradient-free alternative, but the technical mechanism has not been disclosed.

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 goal of the VKORE software layer that VIDRAFT is building is to become "the CUDA of the NPU field," covering three capabilities: inference acceleration, full pretraining, and LoRA fine-tuning; what is released this time is a publicly verified artifact of the pretraining milestone.