Korean AI startup VIDRAFT's new inference accelerator lets users verify speed claims on their own hardware — a bold move in a crowded market.
TL;DR: VIDRAFT, a Seoul-based AI startup, has publicly released a performance benchmark and deployment container for its large language model inference acceleration engine, VKAE. The engine delivers up to 23.4 times higher inference throughput than standard services under identical GPU conditions. Users can reproduce the results independently by running the publicly available container on their own hardware.
VIDRAFT, the Korean AI startup targeting True-AGI by 2030, made a notable move in early July 2026 when it publicly released both a performance benchmark and a ready-to-deploy container for its new kernel-level inference acceleration engine, VKAE — short for VIDRAFT Kernel Acceleration Engine. The announcement marks one of the company's most concrete technical disclosures to date, putting verifiable performance figures in the hands of developers and enterprise users alike.
At the heart of the release is VKAE, a kernel-level acceleration tool designed to dramatically speed up large language model (LLM) inference without sacrificing output quality. According to VIDRAFT, the engine achieves up to 23.4 times higher inference throughput compared to standard serving setups, assuming the same GPU hardware and operating conditions.
To make those numbers independently verifiable, VIDRAFT packaged the model weights and service environment together into a single Docker container that any developer can pull and run on a compatible GPU. This approach reflects what VIDRAFT CEO Kim Min-sik described as a core principle of the release: speed claims without reproducibility are meaningless, and the only credible benchmark is one a user can run themselves.
On the raw performance side, VIDRAFT shared figures from its internal measurements using an NVIDIA B200 GPU at FP8 precision. With its representative model, Qwen3.5-35B-A3B, VKAE generates up to 601 tokens per second in single-stream mode. When handling multiple concurrent requests, throughput surpasses 10,000 tokens per second. VIDRAFT has also applied the same family of optimization techniques to its ultra-large JGOS-398B model, extending the acceleration benefits beyond mid-sized deployments.
The internal mechanics of the kernel remain proprietary and undisclosed, but the public benchmark and unified container provide a legitimate external verification pathway — meaning developers don't need access to VIDRAFT's internal codebase to confirm the performance claims for themselves.
VKAE also ships with an OpenAI-compatible API, which lowers the integration barrier significantly. Teams that already run OpenAI-based pipelines can connect VKAE without overhauling their existing service infrastructure.
In terms of hardware support, VKAE currently targets NVIDIA's Blackwell (B200) and Hopper (H100, H200) GPU families, with support for smaller form-factor cards like the A10G actively being extended.
VIDRAFT framed the release around four pillars: verifiability (users can reproduce benchmark data themselves), cost efficiency (higher throughput means fewer GPUs needed per workload), quality preservation (speed improvements do not degrade model outputs), and ease of integration (OpenAI-compatible API from day one).
The LLM inference optimization space has grown increasingly crowded, with numerous frameworks and vendors competing on throughput benchmarks. What sets VIDRAFT's approach apart — at least in intent — is the emphasis on third-party reproducibility. By shipping a self-contained container rather than asking users to trust published numbers, the Korean AI startup is making a direct challenge to benchmark skepticism that has become common in the industry.
For enterprise buyers, the practical upside is straightforward: if VKAE delivers even a fraction of its headline throughput gains in production, the cost per token drops substantially. That translates into real savings at scale, particularly for teams running high-volume inference workloads.
VKAE also fits into a broader, increasingly coherent product stack that VIDRAFT has been assembling since its founding in 2024. The company has previously released FINAL Bench, a benchmark for measuring AI metacognitive abilities; MARL, a runtime middleware aimed at reducing hallucinations; and a series of proprietary models under the Darwin, Chimera, and Aether names. VKAE slots in at the infrastructure layer, giving VIDRAFT coverage from model development all the way through to serving optimization — an unusually full stack for a two-year-old startup.
The company is headquartered at Seoul AI Hub and has publicly stated a goal of achieving True-AGI by 2030, alongside ongoing work on a Korean-specific LLM and a scientific reasoning model.
Q: What is VIDRAFT's VKAE engine?
A: VKAE, or VIDRAFT Kernel Acceleration Engine, is a kernel-level tool that accelerates large language model inference. It delivers up to 23.4 times higher throughput than standard serving configurations under identical GPU conditions while preserving output quality.
Q: How can developers verify VKAE's performance claims?
A: VIDRAFT has released a publicly available Docker container that packages model weights and the service environment together. Developers can run this container on a compatible GPU — such as an NVIDIA B200, H100, or H200 — to reproduce the benchmark results independently.
Q: Does VKAE integrate with existing OpenAI-based services?
A: Yes. VKAE ships with an OpenAI-compatible API, enabling teams that already use OpenAI-based pipelines to connect the engine without redesigning their service architecture.
Source: Wedoany (영문) (2026-07-07) — original article