Korean AI Startup Vidraft Wins Global Top Prize in Google's Gemma Challenge
Translated from Korean, summarized and contextualized by DistantNews.
At a glance
- South Korean AI startup Vidraft has secured the global first place in the "verified records" category of the "The Fast Gemma Challenge," hosted by Google.
- The competition focused on maximizing the inference speed of AI models within identical environments.
- Vidraft's achievement demonstrates the technical capabilities of Korean startups on the global stage and has implications for cost-effective AI services.
South Korean deep tech company Vidraft has achieved global recognition, clinching first place in the "verified records" category of "The Fast Gemma Challenge." This competition, jointly organized by Google's Gemma team and Hugging Face, highlights Vidraft's prowess in optimizing AI model performance.
The challenge centered on maximizing the inference speed of AI models, specifically Google's "gemma-4-E4B-it" model, using NVIDIA A10G GPUs. Participants were tasked with enhancing performance solely through software optimization. Vidraft's AI agent, "vidraft-darwin," successfully achieved an impressive 510.58 tokens per second (TPS) with a PPL of 2.3929 after rigorous verification by the organizers.
This performance represents a more than six-fold increase in inference speed compared to the original model, while maintaining the quality of responses. Such an improvement directly addresses the critical challenge of cost-efficiency in AI services. Vidraft's achievement means that the same service can potentially be delivered using significantly less infrastructure, a sixth of the original requirement.
Vidraft attributes this success to its proprietary VK inference engine and the optimization technology within its "POCKET" on-device platform. CEO Kim Min-sik emphasized that fast and economical service delivery on identical hardware is key to future AI competitiveness. The company also highlighted its ongoing development in AI infrastructure and models, including its "AETHER" foundation model and the "Darwin Family" framework for enhancing LLM inference without additional training.
Originally published by Dong-A Ilbo in Korean. Translated, summarized, and contextualized by our editorial team with added local perspective. Read our editorial standards.