Korean AI startup VIDRAFT makes its inference acceleration claims independently verifiable — a rare move in a field crowded with unaudited benchmarks.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released its LLM inference acceleration engine called VKAE (VIDRAFT Kernel Acceleration Engine), along with a performance leaderboard and a unified Docker container that lets anyone reproduce benchmark results on their own hardware. The engine delivers up to 23.4× higher throughput compared to standard serving methods on identical GPU hardware. The same optimization technology has been applied to VIDRAFT's flagship large-scale model, JGOS-398B.
VIDRAFT, the Seoul-based Pre-AGI startup led by CEO Minsik Kim, announced on July 8, 2026 the public launch of VKAE — its proprietary kernel-level LLM inference acceleration engine — alongside an open performance leaderboard and a prebuilt, all-in-one Docker container designed to put its speed claims directly in users' hands.
At the center of the announcement is VKAE (VIDRAFT Kernel Acceleration Engine), a kernel-level inference engine built to dramatically reduce the cost and latency of running large language models. According to VIDRAFT, VKAE achieves up to 23.4× greater throughput than conventional standard serving approaches, all within the same GPU environment and without degrading response quality.
What makes this release strategically distinct is not just the engine itself, but the verification infrastructure surrounding it. VIDRAFT is simultaneously releasing two companion tools:
The container also supports OpenAI-compatible APIs, meaning teams can integrate it directly into existing applications and services without rewriting their stack. VIDRAFT has stated that VKAE's optimization methodology has been applied consistently to its largest model, JGOS-398B, demonstrating that the approach scales to frontier-class model sizes.
On the hardware side, VKAE currently offers primary support for NVIDIA's Blackwell (B200) and Hopper (H100, H200) GPU families, with the company actively expanding compatibility to additional hardware configurations, including smaller form-factor options such as the A10G.
VIDRAFT described four pillars that define VKAE's value proposition: transparent, fraud-resistant verification; dramatically lower per-token operational costs; quality stability that ensures output integrity is preserved even at higher throughput; and frictionless integration with systems already in production.
While the underlying kernel implementation remains proprietary — VIDRAFT cites trade secrets as the reason — the company's decision to expose the leaderboard and a fully reproducible container represents an uncommon commitment to third-party auditability in the AI inference space.
CEO Minsik Kim articulated the philosophy directly: "A speed race that third parties cannot reproduce themselves is meaningless in the market. The core of this release is that we've simultaneously launched both the leaderboard and the container, so users can verify performance on their own hardware."
Inference efficiency has rapidly become one of the most commercially consequential battlegrounds in AI. As organizations scale LLM deployments, the cost of serving — measured in tokens processed per second per dollar — can determine whether a product is economically viable. Claims of superior inference speed are common; independently verifiable proof of those claims is not.
VIDRAFT's approach of coupling a benchmark leaderboard with a publicly runnable container directly addresses a credibility problem that has quietly plagued the industry. If a third party can pull a container, run it on hardware they control, and arrive at the same figures, the numbers carry real weight.
This launch also marks a significant expansion of VIDRAFT's publicly visible technical stack. The company has previously introduced the FINAL benchmark for measuring AI metacognition, the MARL runtime middleware for hallucination reduction, and its Darwin, Chimera, and Aether model families. VKAE now completes what the company describes as a full-stack capability — spanning model development, evaluation, reliability validation, and serving optimization under one roof.
Founded at Seoul AI Hub in 2024, VIDRAFT has publicly stated a goal of achieving AGI by 2030. The company is positioning its combination of Korean-language-specialized LLMs, scientific reasoning models, and infrastructure optimization technology as its entry point into the global AI market.
Q: What is VKAE and what does it do?
A: VKAE, or VIDRAFT Kernel Acceleration Engine, is a kernel-level inference engine that accelerates LLM serving. VIDRAFT claims it delivers up to 23.4× higher throughput than standard serving methods on the same GPU hardware while maintaining output quality.
Q: How can users verify VIDRAFT's performance claims?
A: VIDRAFT has released a public performance leaderboard alongside a unified Docker container that bundles model weights and an optimized serving environment. Users can run the container on their own GPU hardware to independently reproduce the benchmark results.
Q: Which hardware does VKAE currently support?
A: VKAE primarily supports NVIDIA's Blackwell (B200) and Hopper (H100, H200) GPU families. VIDRAFT says it is continuing to expand its supported hardware lineup to include additional configurations.
Source: AI타임스 (2026-07-08) — original article