Domestic artificial intelligence (AI) deep-tech company Vidraft announced on the 3rd that it ranked first globally in the verified record category at “The Fast Gemma Challenge,” co-hosted by Google’s Gemma team and Hugging Face. The result firmly demonstrated the technological capabilities of a Korean startup on the global stage.

This competition is designed to maximize the inference speed of AI models under identical conditions. Participants were required to improve performance solely through software optimization using Google’s multimodal model “gemma-4-E4B-it” and an NVIDIA A10G GPU.

The rankings were determined exclusively based on “VERIFIED” records, for which the organizers directly conducted closed tests to confirm speed and quality. Vidraft’s autonomous AI agent “vidraft-darwin” achieved 510.58 tokens per second (TPS) and a PPL of 2.3929 after rigorous verification. This represents an inference speed more than six times faster than the original model while fully maintaining answer quality.

This achievement is directly linked to cost efficiency, a core challenge for AI services. Increasing inference speed by a factor of six means that the same service can be provided with infrastructure at roughly one-sixth the previous level. Vidraft’s VK inference engine and its on-device platform “POCKET” underpinned this performance through optimization technologies.

Vidraft CEO Kim Min-sik emphasized that providing fast and economical services on identical hardware will be the key to future AI competitiveness, and attached significance to the fact that the company’s technology has been objectively verified on the international stage.

Meanwhile, Vidraft is a deep-tech company that covers AI infrastructure and model development, including the foundation model “AETHER.” Recently, it announced technological milestones by unveiling “Darwin Family (model-merging AI framework),” which boosts LLM inference performance without additional training, and “MARL,” a runtime middleware that reduces hallucinations.

Popular News