Korea's Pre-AGI startup makes its speed claims independently verifiable for the first time.
TL;DR: VIDRAFT, a Seoul-based AI startup, has publicly released a performance leaderboard and a unified Docker container for its LLM inference acceleration engine, VKAE (VIDRAFT Kernel Acceleration Engine). The engine has recorded up to 23.4× higher inference throughput than standard serving under identical GPU conditions. Users can now reproduce benchmark results directly on their own hardware without relying solely on VIDRAFT's internal claims.
VIDRAFT, the Korean Pre-AGI startup headquartered at Seoul AI Hub, took a significant step toward transparent AI benchmarking on July 6, 2026, by unveiling a public performance leaderboard and a distributable Docker container for its kernel-level LLM inference engine, VKAE. The move is designed to let engineers and enterprises verify the engine's speed claims independently — an increasingly rare commitment in an industry often accused of unverifiable performance marketing.
VKAE, short for VIDRAFT Kernel Acceleration Engine, is a kernel-level acceleration layer built to dramatically increase the throughput and responsiveness of large language model (LLM) inference. According to VIDRAFT's own measurements, the engine delivers up to 23.4 times the inference throughput of standard serving frameworks when run on the same GPU under identical conditions — while preserving output quality.
The two centerpiece releases accompanying this announcement are:
A public performance leaderboard. VIDRAFT has opened a leaderboard where benchmark results for supported models are displayed, giving the broader developer community a reference point for comparing VKAE's performance across different configurations.
A unified Docker container. Rather than asking users to trust published figures alone, VIDRAFT is distributing a pre-packaged container that bundles model weights and the serving environment together. Anyone with a compatible GPU can pull the container and run the benchmarks themselves. The container also exposes an OpenAI-compatible API, meaning teams can plug VKAE directly into existing applications with minimal integration effort.
VIDRAFT's own measurements on an NVIDIA B200 GPU in an FP8 environment showed that the Qwen3.5-35B-A3B model produced up to 601 tokens per second in a single-stream scenario. Under multi-request conditions, throughput exceeded 10,000 tokens per second. The same family of optimization techniques has been applied to JGOS-398B, VIDRAFT's very large model. The company has made clear that the internal kernel implementation is proprietary and will not be disclosed, but the leaderboard and container are intended to provide a credible external validation path.
GPU support currently centers on NVIDIA's Blackwell (B200) and Hopper (H100, H200) architectures, with the company actively expanding coverage to include smaller cards such as the A10G.
CEO Min-sik Kim framed the release around accountability: "A speed race that cannot be reproduced is meaningless. The core of VKAE is that users can verify performance directly on their own hardware," he said, as reported by 디지털데일리.
The LLM inference optimization space is crowded, and benchmark credibility has become a genuine problem. Vendors routinely publish throughput figures that are difficult or impossible to reproduce outside of tightly controlled internal environments. By shipping a self-contained Docker container alongside a live leaderboard, VIDRAFT is effectively inviting public scrutiny — a posture that distinguishes it from many competitors.
VIDRAFT identifies four concrete benefits it believes VKAE delivers: independent reproducibility of published numbers, meaningful reductions in serving costs that follow directly from higher throughput, preservation of model output quality even as speed increases, and rapid integration enabled by OpenAI API compatibility. Each of these addresses a common objection enterprises raise when evaluating third-party inference engines.
The announcement also fits into a broader VIDRAFT product narrative. The company has previously released FINAL Bench, a benchmark for measuring AI metacognitive ability; MARL, a runtime middleware aimed at reducing hallucinations; and multiple model families under the Darwin, Chimera, and Aither names. VKAE represents the infrastructure end of what VIDRAFT describes as a full-stack AI capability — spanning model development all the way through to optimized serving.
Founded in 2024 with an explicit goal of achieving what it calls "True-AGI" before 2030, VIDRAFT has positioned itself as a global-market contender building Korean-language-specialized LLMs, scientific reasoning models, and now publicly auditable inference infrastructure.
Q: What is VKAE and what does it do?
A: VKAE (VIDRAFT Kernel Acceleration Engine) is a kernel-level inference acceleration engine for large language models. It is designed to increase throughput and reduce response latency compared with standard serving frameworks, without degrading output quality.
Q: How can developers verify VKAE's performance claims?
A: VIDRAFT has published a performance leaderboard and is distributing a Docker container that bundles model weights with the serving environment. Developers can run the container on a compatible NVIDIA GPU to reproduce benchmark results on their own hardware.
Q: Which GPUs does VKAE currently support?
A: VKAE's primary support targets NVIDIA Blackwell-generation (B200) and Hopper-generation (H100, H200) GPUs. VIDRAFT has indicated it is expanding support to additional cards, including the A10G.
Source: 디지털데일리 (2026-07-06) — original article