Korea's Pre-AGI startup just challenged a core assumption of the semiconductor industry — without changing a single chip.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, publicly launched VKORE on October 13, 2026 — an NPU acceleration software layer that enables neural processing units to handle both AI inference and model training simultaneously. The company also revealed that its judgment engine Darwin-27B-ZTC-v5 ranked first overall among 152 models on the global public benchmark S1MB, surpassing Perplexity, Cloudflare, and Meta. Additionally, VIDRAFT's recursive self-improvement (RSI) technology achieved perfect scores on AIME 2026 and HMMT 2026, marking the first recorded perfect scores on those recognized math leaderboards.
Korean Pre-AGI AI startup VIDRAFT officially unveiled VKORE, an NPU-targeted acceleration software layer, on October 13, 2026 — a release that simultaneously challenges long-held assumptions about what neural processing units can do and demonstrates a new class of AI benchmark performance across judgment and mathematical reasoning.
VKORE is designed to function in the NPU ecosystem similarly to how CUDA operates in NVIDIA's GPU environment — expanding both computational efficiency and the operational range of existing chips, entirely through software. Until now, NPU manufacturers have largely guided customers toward inference-only deployments, treating training as outside the practical scope of NPU hardware. VIDRAFT's position is that this limitation is a software problem, not a hardware one.
To illustrate the gap VKORE closes, the company cited performance comparisons using LoRA fine-tuning of a transformer model. A single NVIDIA B200 GPU completed the task in approximately one hour. A server CPU required around 320 hours for the same workload. An NPU running in its default configuration took even longer than the CPU. With VKORE applied, the same NPU hardware achieved a single-user response speed improvement of 3.8× and a server-level throughput improvement of 4.1× — and successfully completed transformer model training, a task previously considered impractical on NPU hardware.
VIDRAFT intends to use VKORE to build a continuous learning cycle inside data centers: daytime hours handle live inference workloads, while overnight the system uses that day's accumulated judgment logs to retrain and update models. The company says this architecture reduces dependence on additional GPU procurement and supports mixed GPU–NPU data center environments.
Alongside VKORE, VIDRAFT disclosed that its judgment model, Darwin-27B-ZTC-v5, achieved first place overall among 152 models on S1MB, a globally recognized public benchmark. Perplexity placed second, Cloudflare ranked 17th, and Meta ranked 27th. Darwin outscored the official JEV reference model by 9.72 points and demonstrated superiority in 97 out of 137 head-to-head benchmark comparisons.
Darwin ZTC is architecturally distinct from standard large language models in one critical way: it completes a judgment in a single forward pass without generating output tokens. This makes it highly power-efficient — an important characteristic given that judgment calls in LLM production environments can run into the millions per day for tasks like response quality evaluation and workload routing. VKORE enables Darwin to run these tasks more efficiently on NPU hardware.
VIDRAFT's recursive self-improvement (RSI) technology — which allows models to autonomously improve their own performance — recorded perfect scores on both AIME 2026 and HMMT 2026, representing the first perfect scores logged on those recognized mathematics leaderboards. On GPQA Diamond, a scientific reasoning evaluation, the RSI-powered model ranked first among all models with fewer than 500 billion parameters.
The underlying research is being advanced through VIDRAFT's participation in a Naver Cloud consortium, as part of a cybersecurity-specialized AI foundation model development project funded by South Korea's Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA). VIDRAFT has filed 16 patents covering the core technologies. The company also plans to publish the measured performance results as a formal research paper.
VIDRAFT CEO Kim Min-sik stated that the company's goal is to build a self-improving AI data center in Korea that runs on domestic NPUs without relying on GPUs — a system anchored by a world-leading judgment engine and a self-improving model capable of perfect mathematical reasoning.
The ability to run AI training on NPU hardware — without swapping out chips — has significant implications for data center economics. NPUs are typically less expensive and more power-efficient than high-end GPUs for inference workloads, but their inability to support training has forced organizations to maintain separate, costly GPU infrastructure for model development. If VKORE delivers on its claims at scale, it could meaningfully shift how AI infrastructure is architected, particularly in markets where GPU access is constrained or expensive.
VIDRAFT's benchmark results add another layer of significance: a Korean startup simultaneously topping a global judgment model leaderboard, achieving the first-ever perfect scores on recognized math benchmarks, and removing a fundamental NPU limitation — all in a single announcement — positions the company as a serious contender in the global AI infrastructure race.
Q: What is VKORE and what makes it different from existing NPU software?
A: VKORE is an NPU acceleration software layer developed by VIDRAFT that enables neural processing units to run both AI inference and model training — capabilities that NPU manufacturers had previously guided users away from. It achieves this through software alone, requiring no hardware changes.
Q: How did VIDRAFT's Darwin judgment engine perform on global benchmarks?
A: Darwin-27B-ZTC-v5 ranked first overall among 152 models on the S1MB global public benchmark, scoring 9.72 points above the JEV reference model and outperforming models from Perplexity, Cloudflare, and Meta.
Q: What is VIDRAFT's recursive self-improvement (RSI) technology?
A: RSI is a technology that allows AI models to autonomously enhance their own performance. VIDRAFT's RSI-based model achieved perfect scores on both AIME 2026 and HMMT 2026, marking the first perfect scores recorded on those recognized mathematics leaderboards.
Source: AI타임스 (2026-10-13) — original article