VKORE

VKORE is a software acceleration layer for NPUs, designed to be hardware-agnostic, enabling data centers to mix GPUs and NPUs and still train and serve inference from the same stack.

VKORE itself has not yet been made public; it is currently proprietary infrastructure software in the development stage, with no public SDK, GitHub repository, or API, and its training capability has only been demonstrated on internal NPU hardware.

VIDRAFT plans to use VKORE for two types of internal workloads: the AI judgment engine (which outputs probabilities or scores rather than free text) and the recursive self-improvement model (RSI, in which the model repeatedly solves problems and feeds verified results back into its own training loop).

All of the above performance figures come from VIDRAFT's own public announcements, and there has been no third-party independent reproduction.

According to ZDNet Korea, VIDRAFT officially announced: on the same NPU, after enabling VKORE, single-user inference latency improved 3.8x, concurrent server throughput improved 4.1x, and training throughput was about 170x compared with a CPU baseline.

VKORE's implementation approach is to recombine, in software, the training compute primitives missing from inference-only NPUs (such as gradient computation and weight updates) based on the low-level operations the NPU itself supports, and to maximize hardware utilization through scheduling.

VIDRAFT has filed 16 patents related to VKORE.

VKORE's functional positioning is analogous to what CUDA is to NVIDIA GPUs, sitting between application-layer ML frameworks and NPU hardware, but the specific implementation details (whether it includes a compiler, kernel libraries, runtime scheduler) have not been disclosed.

Under VKORE acceleration, the single-request response latency of the same NPU is 3.8 times faster than the baseline NPU, and the server-level concurrent throughput is 4.1 times higher.

VIDRAFT is a South Korean AI startup that released the VKORE software acceleration layer.

As of the time of reporting, VKORE, Darwin-27B-ZTC-v5 and the RSI engine were all not publicly available, with no Hugging Face repository, GitHub code repository or public API.

VIDRAFT's CEO stated that the goal is to make VKORE run on South Korean domestic NPUs, but the list of supported NPU vendors and chip series has not been made public.

Currently all performance data for VKORE and Darwin-27B-ZTC comes from a single source (a report by AI타임스), lacking independent third-party verification.

VKORE increases single-user inference response speed by 3.8x on the same NPU hardware.

An NPU equipped with VKORE reaches about 75% of a single B200's pretraining throughput.

VKORE uses a software layer to fill the gap in NPUs' training capability, enabling NPUs to execute training workloads without changing their silicon design.

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.

An NPU equipped with VKORE reaches 170% of a single H100's per-device pretraining throughput.

VKORE's numerical behavior during training on NPUs is consistent with standard training; it is real training, not an approximation.

VIDRAFT's VKORE software acceleration layer, without changing the chip, both increases NPU inference performance several times over and makes Transformer LoRA training run on this kind of inference-specific hardware, with numerical precision aligned with standard PyTorch at 0.9999.