VIDRAFT's VKORE: An NPU Acceleration Software Layer That Unlocks Both Inference and Training on Non-GPU Hardware
TL;DR: VIDRAFT has released VKORE, a software acceleration layer for NPUs that — analogous to CUDA for NVIDIA GPUs — dramatically improves inference throughput and, for the first time on these chips, enables transformer model training without any hardware changes. Paired with a top-ranked judge model and a recursive self-improvement (RSI) engine hitting perfect scores on public math benchmarks, VKORE is designed to let data centers run a full inference-by-day / training-by-night loop entirely on NPU silicon. Developers in ML infrastructure and LLMOps should pay attention.
What it is
VKORE is a software acceleration stack built by VIDRAFT (비드래프트), a Korean Pre-AGI AI startup, targeting Neural Processing Units (NPUs) — the class of inference-optimized accelerators that are increasingly common in data centers alongside traditional GPUs.
The core pitch: NPUs have historically been considered inference-only devices. Manufacturers themselves have positioned them that way. VKORE challenges that assumption by operating purely at the software layer — no chip swap required — to unlock:
- High-throughput inference with dramatically lower latency per request
- Transformer model training (specifically LoRA fine-tuning, as tested), which was previously impractical or effectively impossible on NPUs at competitive speeds
VKORE is positioned as the NPU equivalent of what CUDA is to NVIDIA GPUs: a foundational software layer that exposes and maximizes the hardware's computational potential for AI workloads.
How it works
At a conceptual level, VKORE introduces an optimized software execution layer between the ML framework and the NPU hardware. The idea is analogous to a custom kernel/runtime that schedules and batches operations in a way that the default NPU drivers do not.
Without VKORE, running a server CPU on a LoRA training task that takes ~1 hour on an NVIDIA B200 GPU requires roughly 320 hours — and a naively-driven NPU performs worse than that CPU baseline. VKORE restructures how operations are dispatched to the NPU, recovering massive efficiency gains without modifying the underlying silicon.
The intended operational pattern VIDRAFT describes is a day/night inference-training loop:
- Daytime: The NPU cluster serves live inference traffic
- Nighttime: VKORE enables the same NPU cluster to train on that day's accumulated judge logs and feedback signals
This cyclical architecture is designed to reduce dependence on additional GPU procurement and to support hybrid data center environments where GPUs and NPUs coexist.
VKORE also serves as the execution platform for two other VIDRAFT AI components:
- Darwin-27B-ZTC-v5, a judge/decision model
- An RSI (Recursive Self-Improvement) reasoning engine
Benchmarks & results
All figures below are sourced directly from the AI타임스 report:
VKORE inference & training gains (same NPU, VKORE vs. baseline NPU):
- Single-request response latency: 3.8× faster
- Server-level concurrent throughput: 4.1× higher
- LoRA transformer training: successfully completed (previously impractical on NPU)
Darwin-27B-ZTC-v5 (judge model) — S1MB global benchmark:
- Ranked #1 out of 152 models in the S1MB public benchmark
- Outscored the official JEV reference model by 9.72 points
- In 137 head-to-head benchmark comparisons: won 97 out of 137
- Notable rankings of competitors in the same benchmark: Perplexity (#2), Cloudflare (#17), Meta (#27)
Darwin-27B-ZTC uses a zero-token-cost (ZTC) inference design: it completes a judgment via a single forward pass without generating output tokens, making it highly power-efficient for high-frequency operations like response quality scoring and task routing (where millions of calls per day are typical in production LLM systems).
RSI reasoning engine — math benchmarks:
- AIME 2026: 100% (perfect score)
- HMMT 2026: 100% (perfect score) — reported as the first perfect score on a certified math leaderboard
- GPQA Diamond (science reasoning): #1 among models with fewer than 500B parameters
VIDRAFT has stated that full performance results will be released as a research paper.
How to try it
VKORE, Darwin-27B-ZTC-v5, and the RSI engine are not yet publicly available via Hugging Face, GitHub, or a public API as of this report. VIDRAFT has announced the technology and indicated a forthcoming paper, but no public access channel has been disclosed.
Developers interested in access or early partnerships should monitor VIDRAFT's official channels. The company is participating in a government-backed AI foundation model program (under MSIT/NIPA) as part of the NAVER Cloud consortium, which may influence future release timelines.
FAQ
Q: Is VKORE hardware-specific, or can it work on any NPU?
A: The source does not specify supported NPU vendors or chip families. VIDRAFT's CEO stated the goal is to run workloads on domestic (Korean) NPUs, but no supported hardware list has been publicly released. Check VIDRAFT's forthcoming paper for details.
Q: Darwin-27B-ZTC uses a single forward pass — how does that work for judgment tasks?
A: Rather than generating a chain of tokens to arrive at a verdict, Darwin-27B-ZTC is designed as a classification/scoring model that outputs a judgment signal directly from one forward pass. This is architecturally similar to reward models or cross-encoders used in RLHF pipelines, but optimized specifically for high-throughput, low-latency decision-making in production LLM routing and quality-evaluation scenarios.
Q: When will the VKORE paper be available?
A: VIDRAFT has confirmed a paper is planned but no publication date or venue has been announced. The AI타임스 article from October 13, 2026 is the primary public source at this time.
Originally reported by AI타임스 (2026-10-13) — source article.
Top comments (0)