Korean Pre-AGI AI startup VIDRAFT has publicly released VKAE — short for Vidraft Kernel Acceleration Engine — a purpose-built inference acceleration engine designed to dramatically increase the throughput of large language models without degrading response quality. The announcement was covered by Italian AI media outlet iHAL on July 9, 2026.
Rather than applying optimizations at the model or application layer, VKAE operates at the kernel level — the layer closest to hardware execution — where the low-level operations involved in token generation actually take place. According to figures shared by the company, VKAE can increase processing throughput by up to 23.4 times compared with standard serving methods under equivalent GPU conditions.
The practical significance of that figure is considerable. Inference has become one of the dominant cost centers in generative AI deployment. Each request to a language model consumes memory, GPU time, data transfers, and sequentially generated tokens. Compressing the per-token operational cost makes it meaningfully more economical to run chatbots, AI agents, coding assistants, enterprise tools, and any application that must handle large numbers of concurrent users.
VIDRAFT is distributing VKAE as a self-contained Docker container, a choice that carries its own practical weight. Developers and organizations do not need to manually reconstruct a runtime environment, resolve complex dependency chains, or configure individual serving components. The model weights, serving environment, and kernel optimizations are bundled into a single executable package, enabling direct performance testing on an operator's own GPU hardware.
This shifts the conversation from declared benchmarks to reproducible results on real infrastructure — a meaningful distinction in a market where performance claims are often difficult to independently verify.
VKAE also ships with OpenAI-compatible API support, which lowers the barrier to adoption significantly. Many existing AI applications are already built around that API format, meaning VKAE can be slotted into live production pipelines without requiring a full rewrite of the application layer. For enterprises, this translates into reduced integration costs and the ability to test acceleration gains on already-operational workflows.
Notably, VIDRAFT has validated VKAE's optimizations against JGOS-398B, a very large-scale model. This signals that the engine's target scope extends well beyond compact or mid-sized models. At the scale of hundreds of billions of parameters, factors such as memory management, parallelism, and latency become critical bottlenecks — and even incremental kernel-level improvements can produce outsized gains across an entire cluster.
VIDRAFT has chosen to keep the internal kernel implementation proprietary, but has published both a public leaderboard and the integrated container to enable external verification. The approach represents a deliberate balance: the most sensitive implementation details remain protected, while users retain the ability to measure actual system behavior on their own hardware. In an environment where AI performance claims frequently lack reproducibility, this stance positions VKAE as a technically credible offering.
The release reflects a broader industry pressure point: as generative AI scales up, the economics of inference are becoming just as strategically important as model capability itself. VIDRAFT's bet is that kernel-level efficiency, delivered in a reproducible and integration-friendly package, is a viable path to addressing that challenge.
Source: iHAL (이탈리아) (2026-07-09) — original article