The Korean AI startup's new unified container lets developers benchmark kernel-level acceleration on their own hardware.
TL;DR: VIDRAFT, a Seoul-based Korean AI startup, has publicly released a performance leaderboard and a unified Docker container for its in-house LLM inference acceleration engine, VKAE (VIDRAFT Kernel Acceleration Engine). Under identical GPU conditions, VKAE has recorded up to 23.4× higher inference throughput than standard serving frameworks. Users can now reproduce those results directly on their own hardware without complex setup, using an OpenAI-compatible API.
Korean AI startup VIDRAFT announced on July 6, 2026, the public release of a performance leaderboard and an integrated Docker container for VKAE — its proprietary kernel-level engine designed to dramatically accelerate large language model (LLM) inference. The move marks a significant step in the company's push to make its inference technology independently verifiable by the broader developer community.
VIDRAFT has made two things available simultaneously: a public leaderboard documenting VKAE's benchmark results, and a unified Docker container that bundles model weights together with an optimized serving environment. Developers can pull the container and run it on their own GPU hardware to reproduce the published numbers without navigating a complicated installation process.
According to VIDRAFT's own measurements, running the Qwen3.5-35B-A3B model on an NVIDIA B200 GPU in an FP8 environment, VKAE achieved a single-stream peak of up to 601 tokens per second, and exceeded 10,000 tokens per second when handling multiple concurrent requests. Compared with a standard serving stack on identical hardware, that represents up to a 23.4× improvement in throughput. The company states that output quality is preserved throughout — speed gains do not come at the cost of degraded model responses.
The same family of optimization techniques applied to Qwen3.5-35B-A3B has also been extended to VIDRAFT's much larger JGOS-398B model. While VKAE's internal kernel implementation remains proprietary — the company describes it as a core trade secret — the leaderboard and container together give outside users a transparent, reproducible window into real-world performance.
The container supports an OpenAI-compatible API, meaning teams that already have applications built on the OpenAI interface can integrate VKAE-powered serving with minimal code changes. GPU support is currently centered on NVIDIA's Blackwell (B200) and Hopper (H100, H200) families, with the company actively working to expand coverage to additional hardware configurations, including smaller-footprint options such as the A10G.
VIDRAFT's CEO Kim Min-sik was quoted in the original 디지털타임스 report as stating that a speed race that cannot be reproduced is meaningless, and that making the leaderboard and container publicly available so users can verify results on their own hardware is the central point of this release. He added that VIDRAFT is developing not only larger models but also the infrastructure technology to operate large-scale models faster and at lower cost, and that VKAE represents an important pillar of the company's full-stack AI strategy spanning model development through serving optimization.
The decision to pair a performance claim with a self-contained, runnable artifact addresses one of the most persistent credibility problems in AI benchmarking: numbers that cannot be independently checked. By distributing a Docker container alongside the leaderboard, VIDRAFT shifts the burden of proof from "trust our internal results" to "run it yourself."
For enterprise teams evaluating inference infrastructure, the practical upside is meaningful. Higher throughput on the same GPU hardware translates directly into lower per-token serving costs — a factor that becomes decisive at production scale. An OpenAI-compatible API further lowers the switching cost for organizations already invested in that ecosystem.
This release also fits into a broader picture of VIDRAFT's publicly available work. The company has previously released FINAL Bench, a benchmark designed to measure AI metacognition; MARL, a runtime middleware aimed at reducing hallucinations; and its Darwin, Chimera, and Aether model families. VKAE extends that portfolio into the serving and infrastructure layer, completing what VIDRAFT describes as a full-stack approach covering model development, evaluation, reliability, and deployment optimization.
Founded in 2024 and based at the Seoul AI Hub, VIDRAFT has set a goal of achieving True-AGI before 2030. Its current focus spans Korean-language LLMs, scientific reasoning models, inference acceleration infrastructure, and domain-specific AI applications in areas such as drug discovery, materials science, and enterprise productivity.
Q: What is VKAE and who makes it?
A: VKAE (VIDRAFT Kernel Acceleration Engine) is a kernel-level LLM inference acceleration engine developed by VIDRAFT, a Korean AI startup founded in Seoul in 2024. It is designed to increase inference throughput and reduce response latency without degrading model output quality.
Q: How can developers verify VKAE's performance claims?
A: VIDRAFT has published a performance leaderboard alongside a unified Docker container that bundles model weights and the optimized serving environment. Developers can run the container on their own compatible GPU hardware to reproduce the benchmark figures directly.
Q: Which GPUs and APIs does VKAE support?
A: VKAE currently focuses on NVIDIA Blackwell (B200) and Hopper (H100, H200) GPUs, with support being expanded to additional hardware. It exposes an OpenAI-compatible API, allowing integration with existing applications built on that interface.
Source: 디지털타임스 (2026-07-06) — original article