en.Wedoany.com Reported - VIDRAFT recently announced the public release of a performance benchmark for its large language model inference acceleration engine, VKAE, and deployed a Docker container integrating model weights and service environment. The engine, named VIDRAFT Kernel Acceleration Engine, is a kernel-level acceleration tool.

Under the same GPU and conditions, VKAE achieves up to 23.4 times higher inference throughput compared to standard services. The engine improves throughput and response speed while maintaining output quality. Users can directly verify performance in their own GPU environment by running the publicly available container. VKAE also supports an OpenAI-compatible API, enabling connection to existing services.

According to VIDRAFT's internal measurements, on an NVIDIA B200 GPU with FP8 precision, the representative model Qwen3.5-35B-A3B generates up to 601 tokens per second in single-stream mode. In multi-request processing, throughput exceeds 10,000 tokens per second. The same series of optimization techniques has been applied to the ultra-large model JGOS-398B. The internal implementation of the kernel is a trade secret and is not disclosed, but external verification is possible through the benchmark and unified container.

VIDRAFT highlighted four key points of this release: verifiability allowing users to directly reproduce data; reduced service costs due to throughput improvements; quality preservation during speed enhancements; and immediate integration based on the OpenAI-compatible API. VKAE primarily supports NVIDIA Blackwell (B200) and Hopper (H100, H200) series GPUs, with support being extended to models such as the A10G small.

VIDRAFT has previously released the FINAL Bench for measuring AI metacognitive abilities, the MARL runtime middleware for reducing hallucinations, and model series including Darwin, Chimera, and Aether, accumulating full-stack AI capabilities from model development to service optimization. VIDRAFT CEO Kim Min-sik stated that speed competition without reproducibility is meaningless, and the core of VKAE lies in enabling users to directly verify performance on their own hardware.

VIDRAFT is a startup founded in 2024 and based at Seoul AI Hub, with the goal of achieving True-AGI by 2030. The company is developing a Korean-specific LLM, a scientific reasoning model, and inference acceleration infrastructure, while advancing global market expansion.