TeamSparta’s 1.8B Model Secures 1st Place in K-AI Leaderboard Inference Performance

AI upskilling specialist TeamSparta (CEO Lee Bang-gyu) announced that its self-developed lightweight language model 'K-AX Spartan 1.8B' achieved solo first place in the highly demanding inference evaluation 'HLE(Ko)' category on the K-AI Leaderboard. The K-AI Leaderboard, operated by the Ministry of Science and ICT and the Korea Institute of AI (NIA), measures language model performance. TeamSparta secured first place in the 'HLE (Human Last Exam)' category, a benchmark developed by the U.S. AI Safety Center and Scale AI, which consists of 2,500 questions created by experts across 100+ fields. The model scored 0.123 points in HLE(Ko), demonstrating 60% higher performance than the second-place model (0.077 points) and outperforming most top-10 models clustered around 0.068-0.077 points. With just 1.8 billion parameters, the model surpassed large-scale models with hundreds to thousands of billions of parameters, achieving results 20 times smaller than the average of the top 10 models (38.7B) and scoring twice as high as a 236B model despite having only 1/131 the parameters. Its memory requirement is minimal at 256MB, enabling high-level inference in restricted environments like CCTV systems or Raspberry Pi devices.