Xenon Releases Hunmin VLM 397B for Autonomous Computer Operation

Xenon, a generative artificial intelligence firm, published its Hunmin VLM 397B model on September 18, 2026, marking a shift toward specialized computer-operating agents. The model utilizes the Qwen 3.5-397B architecture as its foundation, specifically fine-tuned to interpret graphical user interfaces and execute precise inputs.

The development process focused on enhancing screen recognition, allowing the model to identify and interact with specific UI elements like buttons and input fields. This capability represents a functional evolution from standard text-based generative models to agents capable of direct system manipulation.

Performance metrics indicate the model significantly outperforms its base architecture across multiple evaluation frameworks. On the ScreenSpot Pro benchmark, Hunmin VLM 397B achieved an accuracy score of 75.6, securing the second position on the official Hugging Face leaderboard. It remains the only model from a Korean developer currently featured among the 48 entries on that specific leaderboard.

The model demonstrated substantial gains in OS-level task execution during internal testing. On the OSWorld benchmark, which evaluates performance across 360 distinct Linux tasks, the model recorded a score of 70.5, representing a 22.3-point improvement over the base Qwen 3.5-397B. Similar performance gains were observed in the WindowsAgentArena environment, where the model outperformed the base iteration by 9.1 points.

Technical implementation relied on efficient post-training methods to manage the model’s 397 billion parameters. Xenon utilized a low-rank transfer method followed by supervised and reinforcement learning within an FP8 quantization environment. The entire training pipeline was executed using a cluster of eight Nvidia B200 graphics processing units.

Xenon is also distributing an FP8 version of the model, which effectively halves the required storage footprint for deployment. The company provided the full evaluation methodology and conditions alongside the open-source release to ensure reproducibility for researchers. These assets are intended to support the broader development of action-oriented AI agents in enterprise environments.

The model incorporates specific architectural refinements designed to maintain performance while operating in resource-constrained environments. By utilizing FP8 quantization, the engineers reduced the memory overhead typically associated with 397-billion parameter models. This optimization allows for more efficient deployment on hardware clusters that might otherwise struggle with the computational demands of full-precision inference.

The training pipeline involved a multi-stage approach to align the model with computer-operating tasks. Initial transfer learning was followed by supervised fine-tuning on screen-interaction datasets, ensuring the model could map visual inputs to specific coordinate-based actions. Reinforcement learning was then applied to refine the decision-making process during complex, multi-step workflows on the Linux and Windows operating systems.

The strategic focus on computer operation reflects a broader industry trend toward utility-driven model development. Myung Dae-woo, vice president and chief technology officer at Xenon, stated that the competitive landscape is shifting away from raw parameter scaling. He emphasized that the primary objective is now the efficient integration of specialized capabilities into usable, resource-constrained systems, while maintaining existing Korean-language performance metrics.

This transition toward action-oriented AI suggests that the next phase of model development will prioritize agentic workflows over static text generation. By enabling models to interact directly with software interfaces, developers are creating systems that can bridge the gap between high-level intent and low-level execution. The success of Hunmin VLM 397B on standardized benchmarks provides a quantitative baseline for this shift in capability.

The integration of these capabilities into Xenon’s proprietary OneAgent platform serves as a primary application for the new model. By leveraging the browser-handling and screen-recognition features of the Hunmin series, the company aims to automate complex corporate workflows. This approach prioritizes the practical application of large models over the pursuit of general-purpose benchmarks.

Future development will likely focus on refining the stability of these agents during long-horizon tasks. As Xenon continues to iterate on the Hunmin series, the ability to maintain Korean-language proficiency while expanding system-level control remains a core technical differentiator. The industry will monitor how these models perform in real-world, non-synthetic environments as deployment scales.