Generative AI solutions company Xenon said on Thursday it has released as open source Hunmin VLM 397B, an AI model that recognises computer screens and directly operates them to carry out tasks.

Hunmin VLM 397B is a model that strengthens screen recognition and computer control capabilities based on the open model Qwen3.5-397B, which has 397 billion parameters.

It scored 75.6 on ScreenSpot-Pro, which identifies the location of targets on a screen, ranking second on Hugging Face's official leaderboard. On OSWorld, which assesses the ability to perform real computer tasks in a Linux environment, it scored 70.5, 22.3 points higher than the base model.

Xenon said it used eight Nvidia B200 GPUs to post-train the 397-billion-parameter model. It said it conducted supervised fine-tuning and reinforcement learning under FP8 quantisation to add computer control capabilities with limited computing resources.

Across eight Korean-language benchmarks, it kept the performance gap within 2 points of the base model.

Xenon is releasing the original model and an FP8 version that reduces the size to about half under the Apache 2.0 licence, and will also provide its development and evaluation methodology.

Xenon Vice President and CTO Myung Dae-woo (명대우) said, "The centre of gravity in AI model competition is moving away from simply increasing parameter scale to how efficiently needed capabilities are added to large models and how they are advanced to a level that can be used in practice."