Xenon, a generative artificial intelligence solutions company, said on the 18th that it has open-sourced Hunmin VLM 397B, an AI model that recognizes computer screens and operates them directly to carry out tasks. Built on Qwen 3.5-397B, an open AI model with 397 billion parameters, it identifies where to click on a screen — buttons, input fields and other elements — and performs tasks on an actual computer. Xenon released Hunmin 32B last year and Hunmin VLM 235B earlier this year. Hunmin VLM 397B outperformed the base model in screen recognition, beating both the base Qwen 3.5-397B and Qwen-CUA across five benchmarks; on ScreenSpot Pro it scored 75.6 for accuracy, placing second on the official Hugging Face leaderboard, and Xenon said it is the only model developed by a Korean company among the 48 models currently listed. On OSWorld, involving 360 tasks on the Linux operating system, it scored 70.5, or 22.3 points higher than the base model, and on WindowsAgentArena it scored 9.1 points higher. Xenon transferred capabilities using a low-rank method, then conducted supervised learning and reinforcement learning in an FP8 quantization environment, with the entire post-training process using eight Nvidia B200 graphics processing units.