If you've been waiting for an open-source computer-use AI model you can actually run and inspect, this week delivered a big one. On September 18, Korean generative AI company Xenon open-sourced Hunmin VLM 397B — a model that doesn't just answer questions, but looks at a computer screen, figures out where to click, and carries out tasks on a real machine.
01.What is Hunmin VLM 397B?
Hunmin VLM 397B is built on Qwen 3.5-397B, an open AI model with 397 billion parameters, but with its computer-operation capabilities boosted significantly. The model identifies on-screen elements — buttons, input fields, menus — and then performs the actual actions: clicking, typing, navigating. It's the difference between an AI that tells you how to fill a form and one that fills it for you.
This is the third release in the series. Xenon shipped Hunmin 32B last year and Hunmin VLM 235B earlier this year, each version adding more task-execution muscle while keeping its Korean-language performance intact.
02.Why screen recognition is the hard part
Computer use sounds simple until you see how messy real screens are. Every app has its own layout, its own weird icons, its own tiny dropdowns hidden behind hover states. A model has to translate pixels into meaning — "that rectangle is the Submit button" — before it can act. Get that wrong and the agent clicks the wrong thing, sometimes with real consequences.
Across five benchmarks that measure the ability to locate on-screen elements, Hunmin VLM 397B beat both its base model, Qwen 3.5-397B, and Qwen-CUA, a model specialised for computer operation. On ScreenSpot Pro, a demanding screen-recognition benchmark, it scored 75.6 — second place on the official Hugging Face leaderboard, and the only Korean-developed model among the 48 currently listed there.
That recognition ability carried over to real tasks. On OSWorld, which tests 360 tasks on actual Linux systems, the model scored 70.5 — 22.3 points higher than the base model. On WindowsAgentArena, which does the same for Windows, it came in 9.1 points higher.
03.Eight GPUs. That's it.
Here's the thing nobody's talking about: the entire post-training process used eight Nvidia B200 GPUs. Xenon transferred the computer-operating capabilities of an open model using a low-rank method, then ran supervised learning and reinforcement learning in an FP8 quantisation environment. In a field where frontier labs throw thousands of GPUs at a problem, a 397B-parameter model was upgraded on a cluster that a well-funded university department could realistically access.
Honestly, that detail stopped my scrolling more than the leaderboard numbers did. It suggests the gap between "big lab" and "everyone else" in computer use is narrowing faster than most of us assumed.
Xenon's CTO Myung Dae-woo put it well: the centre of gravity in AI competition is shifting away from simply increasing parameters toward "how efficiently you can add the capabilities that large models need and advance them to a level that is actually usable."
04.What you actually get, and what's next
Xenon didn't just dump a model on Hugging Face. It released the original model, an FP8 version that halves the storage requirement, plus the conditions and methods used to evaluate it — so others can verify the claims rather than take them on faith. That last part matters; too many "open" releases skip it.
The company plans to feed these capabilities into OneAgent, its action-oriented AI product, broadening the range of tasks AI can handle in actual corporate work. We broke down how computer use works — the observe-decide-execute loop — in our GPT-6 Astra beginner explainer if you want the concept from scratch. And the Qwen family this model is built on has been busy too — we covered its new omni-modal launch earlier this week.
Key Takeaways
- Hunmin VLM 397B is an open-source, 397B-parameter computer-use model built on Qwen 3.5-397B
- Second place on Hugging Face's ScreenSpot Pro leaderboard (75.6) — the only Korean model in the top 48
- OSWorld score of 70.5 across 360 real Linux tasks, up 22.3 points over the base model
- Post-trained on just eight Nvidia B200 GPUs using low-rank transfer and FP8 quantisation
- Ships with an FP8 version and full evaluation methods — a genuinely open release
Frequently Asked Questions
What is Hunmin VLM 397B?
Hunmin VLM 397B is a 397-billion-parameter open-source computer-use model from Korean generative AI company Xenon, built on Qwen 3.5-397B and open-sourced on September 18. It identifies on-screen elements — buttons, input fields, menus — and then performs the actual actions: clicking, typing and navigating.
How does Hunmin VLM 397B perform on benchmarks?
It scored 75.6 on ScreenSpot Pro, placing second on Hugging Face's official leaderboard and making it the only Korean model among the 48 listed. On OSWorld's 360 real Linux tasks it scored 70.5 — 22.3 points above the base model — and it came in 9.1 points higher on WindowsAgentArena.
How much compute was needed to train it?
The entire post-training run used just eight Nvidia B200 GPUs. Xenon transferred the computer-operating capabilities of the open base model using a low-rank method, then ran supervised learning and reinforcement learning in an FP8 quantisation environment.
What This Means for You
For Indian developers and students, an open computer-use model changes the maths completely. You're no longer tied to one vendor's API pricing in dollars — you can self-host, fine-tune for Indian apps and workflows (think UPI dashboards, ERP screens, government portals), and build automation products on your own terms. For small teams, this is the moment to start experimenting with agents that operate software, not just chat about it. And if computer use is new to you, our GPT-6 Astra — Zero to Hero Guide has a full chapter on it, written for beginners, walking you from your first prompt to building working automations.