South Korea's Pre-AGI startup shatters the assumption that NPUs can't learn — using software alone.
TL;DR: VIDRAFT, a Korean AI startup, has developed an AI acceleration software layer called VKORE that enables inference-only NPUs to perform model training without any hardware changes. On the same NPU, VKORE delivers up to 3.8× faster single-user inference and up to 4.1× greater concurrent server throughput, while achieving roughly 170× the training performance of a CPU in the same server. The company is now applying VKORE to its own judgment engine and recursive self-improvement models.
VIDRAFT, the Korean Pre-AGI AI startup, unveiled a software acceleration layer called VKORE on October 13, 2026 — a technology that unlocks AI model training on NPUs originally designed and sold exclusively for inference workloads, requiring no chip-level modifications whatsoever.
VIDRAFT's VKORE is positioned as an NPU-native software layer conceptually analogous to NVIDIA's CUDA platform for GPUs: a programmable foundation that lets developers run AI compute workloads on NPU hardware. The crucial difference is that VKORE targets chips that were never intended to support training in the first place.
According to the company, applying VKORE to an existing NPU yields measurable gains across multiple dimensions. Single-user inference response speed improves by 3.8×, and the server's ability to handle simultaneous requests improves by 4.1× — all without swapping out any silicon. More strikingly, VIDRAFT says it successfully trained a Transformer model on an NPU whose manufacturer does not officially support training, and that the same NPU running VKORE outpaced the CPU in the same server by approximately 170× on training tasks.
The company has filed 16 patents related to VKORE and is developing it as a hardware-agnostic software layer, explicitly designed not to be locked to any single chip vendor.
VIDRAFT plans to deploy VKORE across two of its flagship AI systems. The first is its proprietary judgment engine — a technology that, rather than generating text, returns a probability score or verdict in response to input data. The second is its Recursive Self-Improvement (RSI) model architecture, in which an AI solves problems autonomously and feeds verified results back into its own training loop.
On the same day as the VKORE announcement, VIDRAFT reported that its judgment engine model Darwin-27B-ZTC-v5 ranked first overall among 152 models on the global public benchmark S1MB. The company also stated that its Darwin-series models achieved perfect scores on both the AIME 2026 and HMMT 2026 mathematics reasoning evaluations.
CEO Min-sik Kim commented: "We have broken the conventional wisdom that NPUs are only for inference — using software alone. We will be the first in Korea to build an AI data center that improves itself, running our world-leading judgment engine and a math-perfect self-improvement model together on domestic NPUs, without depending on GPUs."
The global AI infrastructure conversation has been dominated by GPU supply chains, and for good reason: training large models at scale has historically demanded GPU-class parallel compute. NPUs have carved out a substantial niche in inference — they are power-efficient and purpose-built for forward-pass operations — but they have generally been considered unsuitable for the backward-pass gradient computations that training requires.
If VIDRAFT's claims around VKORE hold up under independent scrutiny, the implications are significant. A software-only path to turning inference NPUs into training-capable hardware would reduce the cost and complexity of building sovereign AI infrastructure — particularly for countries and organizations that want to avoid dependence on a single GPU supplier.
For South Korea specifically, this fits a broader national technology strategy. VIDRAFT's stated long-term goal is to operate a unified AI data center where a single NPU infrastructure handles both inference and continuous model training — a closed-loop system where the AI can improve itself on domestic hardware.
The VKORE announcement also comes at a moment of momentum for VIDRAFT's model portfolio. The company previously claimed top positions across ten Hugging Face benchmark categories and received research-institution recognition on China's equivalent model hub, suggesting the Darwin model family is gaining traction in international evaluation circles.
Q: What is VKORE and what does it do?
A: VKORE is an AI acceleration software layer developed by VIDRAFT that enables NPUs — chips normally limited to inference — to also perform AI model training, using software optimization rather than hardware changes.
Q: How much faster does VKORE make NPU inference?
A: According to VIDRAFT, VKORE improves single-user inference response speed by 3.8× and concurrent server processing capacity by 4.1× on the same NPU hardware.
Q: What benchmark results did VIDRAFT's Darwin model achieve alongside this announcement?
A: VIDRAFT's Darwin-27B-ZTC-v5 placed first overall among 152 models on the S1MB global public benchmark, and Darwin-series models recorded perfect scores on the AIME 2026 and HMMT 2026 mathematics reasoning evaluations.
Source: ZDNet코리아 (2026-10-13) — original article