Xenon Hunmin VLM 397B Is Open, but Qwen-CUA Still Wins the Harder Test

Xenon has released Hunmin VLM 397B as open source, giving a 397-billion-parameter model new computer-control abilities after training on eight Nvidia B200 GPUs. The release makes screen recognition, clicking, typing, and multi-step desktop work available under an Apache 2.0 license. Yet Xenon’s own results reveal an important split between recognizing interface elements and completing longer tasks.

Xenon Hunmin VLM 397B scored 75.6 on ScreenSpot-Pro, a benchmark that measures whether a model can locate interface targets in professional applications. According to the company, that result placed it second on the benchmark’s Hugging Face leaderboard when the model launched. It also beat both its Qwen3.5 base model and Qwen-CUA across five reported interface-grounding evaluations.

The hierarchy changes when the test moves from locating a button to finishing an entire workflow. Xenon reports a 70.5 score on its 360-task OSWorld evaluation, compared with 77.1 for Qwen-CUA under the same reproduced setup. That gap makes this release more revealing than a simple leaderboard announcement.

Hunmin is evidence that a company can transfer a specialized capability into a huge open model with a relatively small training cluster. It is not evidence that visual accuracy alone produces the most reliable computer agent. For developers and enterprise buyers, that distinction matters more than the model’s headline parameter count.

Xenon Hunmin VLM 397B Adds Computer Control to Qwen3.5

The release turns a general multimodal model into an open computer-use model without retraining its entire 397-billion-parameter foundation.

Xenon, also known internationally as GenON, released the model on September 18, 2026. The company publishes its models through the mncai organization on Hugging Face. Its new repository includes the full-precision model, an FP8 version, evaluation results, and implementation guidance.

Hunmin starts with Qwen3.5-397B-A17B, an open-weight mixture-of-experts model from Alibaba’s Qwen team. It has 397 billion total parameters but activates about 17 billion for each forward pass. That design offers the knowledge capacity of a very large model without using every parameter for every token.

The original Qwen3.5 release introduced native multimodal processing for text, images, and video. Xenon retained that foundation while adding computer-use capabilities. These abilities let the model interpret screenshots, produce coordinates, and decide which interface action should happen next.

The model’s full public name is Hunmin-397B-A17B-CUA. CUA means computer-use agent, a system that performs tasks through graphical interfaces used by people. Instead of relying only on a website API, such a model can operate browsers, office applications, and desktop software through visible controls.

This approach expands Xenon’s earlier Hunmin work. The company released Hunmin 32B in 2025 and followed with Hunmin VLM 235B in early 2026. Its company timeline also connects the Hunmin series with OneAgent, an action-oriented enterprise AI product.

Xenon says the new model preserves most of Qwen3.5’s Korean-language performance. Across eight reported Korean benchmarks, its scores remained within approximately two points of the base model. Several individual results rose slightly, while others declined.

That balance is strategically important for Xenon. A computer-use model that loses too much language ability can struggle with instructions, documents, and localized interfaces. Maintaining Korean performance gives the company a clearer enterprise position than a model optimized only for English-language web tasks.

The public model card provides more detail than the launch headlines. It identifies the training data, software stack, hardware configuration, comparison settings, and known evaluation limits. It also labels the baseline results as author-reproduced rather than provider-reported scores.

That disclosure improves the release’s usefulness. Developers can see which comparisons share the same setup, rather than combining incompatible numbers from unrelated leaderboards. It also exposes the tradeoff that defines Hunmin’s current position.

Eight B200 GPUs Carried the Post-Training Work

Xenon’s most consequential claim concerns training efficiency, not the model’s total parameter count.

Training a 397-billion-parameter foundation model from scratch would require far more than eight GPUs. Xenon did not attempt that. It transferred an existing computer-use capability into Qwen3.5, then refined the resulting model through supervised learning and reinforcement learning.

The first step used Qwen-CUA as the source of computer-control behavior. Xenon calculated the parameter differences between Qwen-CUA and its corresponding Qwen3.5 base checkpoint. It then approximated those differences through truncated singular value decomposition, a mathematical method for compressing a large matrix into lower-rank components.

Those compressed differences became LoRA-style adapters. LoRA, or low-rank adaptation, represents a model update with smaller trainable matrices instead of changing every original weight independently. Xenon merged those adapters into the larger Qwen3.5-397B-A17B checkpoint.

This technique matters because conventional fine-tuning becomes difficult as a model grows. Optimizer states, gradients, activations, and model weights all consume GPU memory. Updating a limited low-rank representation reduces the amount of trainable data and the associated memory burden.

Xenon then quantized the transferred checkpoint to FP8. FP8 stores many numerical values with eight bits, reducing memory use compared with 16-bit formats. The team conducted further post-training through FP8 QLoRA, supervised fine-tuning, and agent reinforcement learning.

Supervised fine-tuning taught the model from examples containing desired interface actions. Agent reinforcement learning then rewarded successful behavior inside computer environments. Xenon identifies GRPO, or Group Relative Policy Optimization, as the reinforcement method used during this stage.

The training stack included MS-Swift, Ray, Megatron-Core, and vLLM. Xenon names OpenGVLab’s ScaleCUA-Data as its supervised dataset. It also lists ScaleCUA and CUA-Gym as reinforcement-learning environments.

According to Xenon, the complete post-training process ran on eight Nvidia B200 GPUs. That statement does not mean the full base model was created on eight accelerators. It means the capability-transfer and post-training stages used that cluster after Qwen had already built the foundation model.

This distinction prevents an impressive engineering result from becoming a misleading compute comparison. Xenon benefited from Qwen’s original model, Qwen-CUA’s specialized checkpoint, open training datasets, and established optimization software. Its contribution lies in combining those components efficiently and adapting them for its target capabilities.

The method offers a possible template for smaller AI vendors. Instead of financing a new foundation model, a company can start with an open checkpoint and add a narrow operational skill. It can then spend limited compute on task data, evaluation, and reinforcement rather than repeating general pretraining.

That path also reduces the distance between academic work and deployable enterprise models. A team can focus on applications, local languages, and controlled environments. The resulting model can still inherit broad visual and language abilities from its foundation.

However, the model remains demanding at inference time. A mixture-of-experts architecture activates only part of its parameter set, but all 397 billion weights still require storage. Xenon’s FP8 release approximately halves storage needs, yet deployment still calls for serious multi-GPU infrastructure.

The eight-GPU training claim should therefore not be read as an eight-GPU deployment promise. Training adapters and serving a merged model present different resource requirements. Enterprise teams must evaluate memory capacity, interconnect bandwidth, latency, concurrency, and energy use before choosing this architecture.

Screen Recognition Is Hunmin’s Clearest Advantage

Hunmin’s strongest results show that it can identify interface targets, especially in dense professional screens.

Graphical user interface grounding links a written instruction to a specific location on a screenshot. If a user asks an agent to select a spreadsheet formula, the model must find the relevant cell or toolbar control. A small coordinate error can send the agent into the wrong menu or alter the wrong file.

ScreenSpot-Pro tests that ability across high-resolution professional applications. Its tasks include interfaces that are denser and less forgiving than typical consumer websites. The ScreenSpot-Pro benchmark was designed because earlier grounding tests did not adequately represent specialized desktop software.

Xenon reports a 75.6 result with single-view inference and 76.6 with a zoom-in method. Its reproduced Qwen3.5 base scores were 72.7 and 75.1, respectively. Qwen-CUA scored 62.2 without zoom and 71.9 with zoom in the same reported environment.

Hunmin also scored 96.2 on ScreenSpot-v2, compared with 95.2 for its base model and 95.0 for Qwen-CUA. On OSWorld-G, a grounding-focused evaluation, it reached 79.9. On the refined OSWorld-G-R version, it scored 86.8.

The model card says these grounding results led the comparison group evaluated by Xenon under its reproduced setup. That wording is important. It does not establish permanent leadership across every model, evaluation configuration, or subsequent submission.

The reported results still show a coherent pattern. Hunmin improves interface perception across several tests, rather than relying on a single favorable number. Its largest advantage over Qwen-CUA appears on ScreenSpot-Pro, where the gap reaches 13.4 points under Xenon’s setup.

For enterprise software, this strength has practical value. Dense interfaces remain common in accounting systems, customer relationship platforms, development environments, and internal administration tools. Many of these products offer incomplete APIs or require actions that span visible interface states.

A grounded model can also work with older applications that were never designed for AI integration. It can identify controls from pixels and labels instead of depending on structured webpage elements. That provides broader coverage, although it also creates reliability and security challenges.

Xenon instructs users to provide native-resolution screenshots and normalize coordinates to a range from zero to 1,000. The resulting coordinates must then be scaled to the screen’s actual width and height. These details can materially affect click accuracy.

The model also uses a visual-history limit of five screenshots for its reported computer-use evaluations. Xenon says a history closer to 20 images can increase memory use and inference cost. The smaller window reflects a deployment choice, not only a benchmarking choice.

Limiting visual history can keep serving costs more manageable. It can also leave an agent with less evidence about earlier actions, dialogs, or page states. That tension becomes more important as a task grows longer.

A single accurate click does not require extensive memory. Completing a 30-step workflow does. The difference explains why Hunmin’s strong grounding scores do not automatically translate into leadership on complete computer tasks.

Qwen-CUA Still Leads on End-to-End Computer Tasks

The central result is a split decision: Hunmin sees interface targets better, while Qwen-CUA finishes more long-horizon tasks.

OSWorld evaluates agents inside real computer applications rather than testing isolated coordinates. Tasks can involve browsers, file operations, email clients, image editors, document tools, and multiple applications. The agent must observe a state, choose an action, recover from changes, and continue until completion.

The official OSWorld benchmark originally assembled 369 computer tasks with reproducible initial states and execution-based evaluation. Xenon used a 360-task configuration for the results in its model card. It reports that all compared models used the same visual-history limit of five images.

Hunmin scored 70.5 in that evaluation. The Qwen3.5 base model reached 48.2, making Hunmin’s gain 22.3 points. That is a substantial improvement and supports Xenon’s claim that its post-training added meaningful computer-control behavior.

Yet Qwen-CUA scored 77.1 under the same reproduced setup, leading Hunmin by 6.6 points. WindowsAgentArena produces the same ordering. Hunmin scored 50.9, up 9.1 points from the base model’s 41.8, while Qwen-CUA reached 57.3.

The results do not invalidate Hunmin’s grounding gains. They show that an agent needs more than precise visual targeting. It must plan, track progress, manage errors, understand application state, and avoid actions that derail the workflow.

Consider an agent asked to update figures in a spreadsheet and send the result by email. It might correctly locate every menu, cell, and button. It can still fail by editing the wrong workbook, applying a formula to the wrong range, or attaching an outdated file.

Longer tasks compound small mistakes. An early error changes later screenshots and invalidates the original plan. The model must detect that divergence and recover instead of confidently continuing down the wrong path.

Hunmin’s training results reveal where some improvement came from. Xenon says its capability-transferred checkpoint reached 66.2 on OSWorld before later training. Supervised fine-tuning and reinforcement learning lifted that score to 70.5, a further gain of 4.3 points.

This progression supports the value of Xenon’s method. Capability transfer delivered much of the initial computer-use ability, while task-specific post-training improved it. However, the remaining Qwen-CUA lead suggests some behavior was compressed imperfectly or changed during integration.

The Korean-language objective may also shape the tradeoff. Hunmin preserves performance across eight Korean benchmarks better than Qwen-CUA in several comparisons. Xenon did not optimize only for the highest possible OSWorld score.

For example, Hunmin scored 76.1 on KMMLU-Pro, compared with 71.4 for Qwen-CUA. It also reached 80.3 on K-MMStar, compared with 75.0. Those results suggest a more balanced multilingual profile, although they remain company-reproduced evaluations.

That balance can matter for Korean enterprises. A model must understand local instructions, documents, terminology, and user interfaces before its clicking ability becomes useful. A higher English-centric agent score does not settle every procurement decision.

Still, Xenon should not present grounding leadership as equivalent to overall computer-use leadership. Its model card avoids that mistake by publishing the unfavorable OSWorld comparison. The launch coverage is less useful when it emphasizes second place on ScreenSpot-Pro without explaining the end-to-end gap.

For buyers, the right comparison depends on the intended workload. A workflow dominated by dense interface targeting may favor Hunmin’s profile. A longer sequence requiring recovery and planning may expose its current weakness against Qwen-CUA.

The model’s openness gives teams another option. Developers can inspect the release, adapt it, and run controlled evaluations on their own software. That flexibility is valuable, but it does not remove the need for application-specific testing.

Open Weights Do Not Eliminate Deployment Risk

Apache 2.0 access improves control, but a downloadable computer-use model remains far from a safe autonomous worker.

Xenon released both the original model and an FP8 variant under Apache 2.0. That license generally permits commercial use, modification, and redistribution subject to its conditions. It gives companies more deployment freedom than a closed agent available only through a hosted interface.

Open access also allows private evaluation. An enterprise can test Hunmin against internal applications without sending screenshots to a third-party model provider. Teams can measure its behavior on local language, specialized forms, and organization-specific workflows.

However, model access is only one layer of a computer-use system. A production agent still needs a runtime, action controller, permissions model, audit trail, credential strategy, and recovery process. It also needs defenses against malicious instructions embedded in webpages or documents.

A computer-use model can interact with whatever the operating account can reach. That may include email, customer records, cloud storage, internal dashboards, and financial applications. One mistaken action can have consequences that a wrong chatbot answer does not.

Screen grounding benchmarks provide little evidence about authorization or intent. Locating the correct “Delete” button is technically successful even when pressing it would violate policy. The surrounding agent must decide whether the action is allowed and whether human approval is required.

Benchmark tasks also begin from controlled states. Real desktops accumulate notifications, software updates, browser extensions, expired sessions, custom layouts, and inconsistent window sizes. These variations can change both what the model sees and what an action does.

Xenon acknowledges several evaluation limits. It says results depend on screenshot resolution, reasoning settings, interaction limits, serving software, and benchmark versions. It also warns against comparing its scores directly with results produced using much longer screenshot histories.

That caution should shape how readers interpret every number in the release. The company reproduced baseline models inside its own environment. This creates a fairer internal comparison, but independent teams have not yet broadly replicated the results.

Leaderboard placement presents another uncertainty. ScreenSpot-Pro rankings can change as new models and inference methods appear. A second-place result at launch is a dated snapshot, not a permanent status.

The parameter scale introduces operational risk as well. Even with FP8 quantization and sparse activation, organizations need suitable infrastructure. A smaller specialized model may deliver sufficient accuracy with lower latency, simpler deployment, and easier fine-tuning.

Teams should therefore compare systems at the workflow level. Useful measures include completion rate, intervention frequency, average steps, recovery success, latency, and the severity of failed actions. A single aggregate benchmark score cannot capture all of these dimensions.

The same principle applies to knowledge work surrounding agent operations. Teams need searchable records of instructions, outputs, exceptions, and human decisions. A controlled AI knowledge base can help reviewers reconstruct why an automated task produced a particular result.

This does not make an agent safe by itself. It creates evidence for review, debugging, and policy improvement. Computer-use systems become easier to govern when their actions remain connected to the documents and decisions that authorized them.

Xenon’s release deserves credit for publishing limitations and unfavorable comparisons alongside its strongest scores. That transparency lets developers ask better questions. The next step is independent testing that measures both capability and failure consequences.

Three Signals Will Show Whether Hunmin Can Close the Gap

Hunmin’s significance now depends on independent reproduction, better long-horizon results, and adoption beyond a model repository.

The first signal is independent benchmark replication. External researchers need to run Hunmin, Qwen3.5, and Qwen-CUA under the same hardware, screenshot history, reasoning settings, and task harness. Similar results would strengthen Xenon’s claim that low-rank transfer preserved useful computer-control capabilities.

Replication should cover more than ScreenSpot-Pro. Researchers should test grounding, full task completion, recovery, latency, and resource consumption. They should also publish failure categories, not only aggregate scores.

If independent runs reproduce the 75.6 grounding result and the 70.5 OSWorld score, confidence in Xenon’s evaluation will rise. Large deviations would weaken the launch narrative and point toward configuration sensitivity.

The second signal is whether Xenon narrows the long-horizon gap with Qwen-CUA. Its current model already finds targets accurately, so the next gains must come from planning and state management. Better memory, stronger reinforcement environments, or improved error recovery could raise end-to-end completion.

A future checkpoint should reduce the 6.6-point reported OSWorld deficit without sacrificing Korean performance. That outcome would support Xenon’s argument that efficient post-training can produce a balanced, practical computer-use model. Grounding improvements alone would leave the central limitation unresolved.

The third signal is deployment inside OneAgent or another enterprise product. A public checkpoint demonstrates technical access, but product use reveals whether the model survives latency constraints and unpredictable interfaces. Real adoption would also show whether companies value local-language control enough to accept the model’s infrastructure requirements.

Xenon says it plans to use the new capabilities to improve OneAgent and expand the work AI can perform. Buyers should watch for named deployments, repeatable workflows, intervention rates, and evidence that tasks remain reliable outside benchmark environments.

The most useful case studies will describe boundaries as clearly as successes. They should identify which applications the agent can operate, which actions require approval, and how often a human must intervene. Without those details, customer announcements will offer little technical evidence.

Xenon Hunmin VLM 397B is already a notable engineering release. It shows that a relatively focused team can add computer-control behavior to an enormous open model using eight high-end GPUs for post-training. Its own numbers also prevent an easy victory narrative.

Developers evaluating the model should begin with a narrow workflow and a reversible sandbox. Measure complete task success, not just click accuracy, and record every intervention or recovery. Then compare those results with Qwen-CUA and smaller alternatives under identical conditions. The open license makes that work possible, but only independent testing will establish whether Hunmin’s grounding advantage becomes dependable action.