Korean AI startup VIDRAFT makes verifiable inference acceleration available to anyone with compatible GPU hardware.
TL;DR: VIDRAFT, a South Korean AI startup, has publicly released performance rankings and a unified Docker container for its VKAE (VIDRAFT Kernel Acceleration Engine), a kernel-level LLM inference accelerator. Under the same GPU conditions as standard serving setups, VKAE delivers up to 23.4 times higher inference throughput. The engine is OpenAI API-compatible, allowing immediate integration with existing services.
VIDRAFT, the Seoul-based Pre-AGI startup founded in 2024, has unveiled a major inference acceleration milestone: the public release of performance benchmarks and a deployable Docker container for its VKAE engine, announced on July 7, 2026. The move is designed to let developers and enterprises independently verify the engine's performance claims using their own GPU hardware — no trust required, just a container pull.
The VIDRAFT Kernel Acceleration Engine, or VKAE, is a kernel-level optimization tool built to dramatically increase the throughput of large language model (LLM) inference workloads. Rather than asking the industry to take its numbers at face value, VIDRAFT published an official performance leaderboard and packaged its model weights and serving environment together into a single, publicly accessible Docker container.
The headline figure is striking: under equivalent GPU conditions, VKAE achieves up to 23.4 times the inference throughput of standard LLM serving setups — while preserving output quality throughout the acceleration process.
On a representative benchmark using an FP8 precision configuration, VIDRAFT's own measurements show that the Qwen3.5-35B-A3B model generates up to 601 tokens per second in a single-stream scenario. When handling multiple concurrent requests, throughput surpasses 10,000 tokens per second. The company's ultra-large-scale model, JGOS-398B, applies the same family of optimization techniques.
VKAE's internal kernel implementation remains proprietary — consistent with standard industry practice — but the external verification pathway through the public container and leaderboard means third parties can reproduce and validate results independently in their own environments. GPU support currently centers on NVIDIA's Blackwell series (B200) and Hopper series (H100, H200), with expansion toward smaller form-factor GPUs also underway.
Because VKAE exposes an OpenAI-compatible API, teams already running OpenAI-standard toolchains can connect VKAE to their existing workflows without significant re-engineering.
VIDRAFT CEO Kim Min-sik was quoted as saying that a speed competition that cannot be reproduced is meaningless, and that the core philosophy behind VKAE is enabling users to verify performance directly with their own hardware.
Independent reproducibility has become a critical credibility issue in the AI inference space. Benchmark numbers published in isolation — without a pathway for external validation — are increasingly met with skepticism by enterprise buyers and the developer community alike. VIDRAFT's decision to ship a self-contained Docker container that bundles model weights with the serving environment is a direct response to that skepticism, and it sets a transparent verification standard that few inference optimization vendors have matched publicly.
The business implications are also concrete. A throughput increase of up to 23.4 times on the same GPU hardware translates directly into lower per-token serving costs — a key concern for any organization running LLMs at scale. By increasing the number of requests a single GPU node can handle, operators can defer or reduce hardware expenditure without sacrificing response speed or output quality.
VKAE also fits neatly into VIDRAFT's broader technology stack. The company has previously released FINAL Bench, a metacognition measurement framework for AI systems; MARL, a runtime middleware layer aimed at reducing hallucinations; and the Darwin, Chimera, and Aether model families. Together, these form a vertically integrated, full-stack AI capability spanning model development through to inference optimization. VKAE is the latest — and most externally verifiable — layer in that stack.
Headquartered at Seoul AI Hub, VIDRAFT is pursuing a stated goal of achieving True-AGI by 2030. Alongside VKAE, the company is actively developing a Korean-dedicated LLM and a scientific reasoning model, with global market expansion on its roadmap.
Q: What is VIDRAFT's VKAE engine?
A: VKAE (VIDRAFT Kernel Acceleration Engine) is a kernel-level inference acceleration tool for large language models that increases throughput by up to 23.4 times compared to standard serving setups, while maintaining output quality.
Q: How can developers verify VKAE's performance claims?
A: VIDRAFT has released a public Docker container that packages model weights and the serving environment together, allowing anyone with a compatible NVIDIA GPU to run the container in their own infrastructure and check the results against VIDRAFT's published performance leaderboard.
Q: Which GPU hardware does VKAE currently support?
A: VKAE primarily targets NVIDIA's Blackwell series (B200) and Hopper series (H100, H200) GPUs, with the company working to broaden support to additional GPU models.
Source: Wedoany (중국) (2026-07-07) — original article