Hunmin-397B-A17B-CUA

<style> /* Limit these typography rules to the model card. */ #hunmin-model-card { font-size: 16px; line-height: 1.6; letter-spacing: 0; overflow-wrap: break-word; } #hunmin-model-card h1 { margin: 0 0 12px; padding: 0; font-size: 30px; line-height: 1.25; letter-spacing: -0.018em; } #hunmin-model-card h2 { margin: 26px 0 8px; padding: 0; font-size: 22px; line-height: 1.3; letter-spacing: -0.012em; } #hunmin-model-card h3 { margin: 18px 0 6px; font-size: 18px; line-height: 1.35; letter-spacing: -0.008em; } #hunmin-model-card p { margin: 0 0 10px; line-height: 1.6; letter-spacing: 0; } #hunmin-model-card ul, #hunmin-model-card ol { margin: 6px 0 12px; padding-left: 22px; } #hunmin-model-card li { margin: 4px 0; padding-left: 2px; line-height: 1.55; letter-spacing: 0; } #hunmin-model-card li > p { margin: 0; line-height: inherit; } /* Keep the paragraph immediately above each table close to it. */ #hunmin-model-card > p:has(+ table) { margin-bottom: 6px; } #hunmin-model-card > p:has(> em:only-child) { margin: 6px 0 14px; font-size: 13px; line-height: 1.45; } #hunmin-model-card table { width: 100%; margin: 6px 0 12px; border-collapse: collapse; table-layout: fixed; font-size: 14px; line-height: 1.45; letter-spacing: 0; font-variant-numeric: tabular-nums lining-nums; } #hunmin-model-card th, #hunmin-model-card td { padding: 7px 10px; vertical-align: middle; line-height: 1.45; } #hunmin-model-card th:first-child { width: 43%; } #hunmin-model-card th:not(:first-child), #hunmin-model-card td:not(:first-child) { text-align: right; } #hunmin-model-card td:not(:first-child) { white-space: nowrap; } #hunmin-model-card u { text-underline-offset: 3px; text-decoration-thickness: 1px; } #hunmin-model-card blockquote { margin: 12px 0 16px; padding: 0 0 0 14px; font-size: 15px; line-height: 1.55; } #hunmin-model-card blockquote p { margin: 0; line-height: inherit; } #hunmin-model-card pre { margin: 10px 0 14px; padding: 14px 16px; overflow-x: auto; line-height: 1.5; letter-spacing: 0; } #hunmin-model-card code { font-size: 0.88em; letter-spacing: 0; } #hunmin-model-card :not(pre) > code { white-space: nowrap; } #hunmin-model-card pre code { padding: 0; font-size: 13px; line-height: inherit; } /* 표 소제목: 이전 내용과 분리하고, 아래 표와는 가깝게 */ #hunmin-model-card h3 { margin: 28px 0 8px; padding: 0; font-size: 16px; font-weight: 700; line-height: 1.4; letter-spacing: 0; } /* Benchmarks 바로 아래 첫 소제목은 과하게 벌리지 않기 */ #hunmin-model-card h2 + h3 { margin-top: 16px; } @media (max-width: 640px) { #hunmin-model-card h1 { font-size: 26px; } #hunmin-model-card h2 { margin-top: 24px; font-size: 21px; } #hunmin-model-card table { font-size: 13px; } #hunmin-model-card th, #hunmin-model-card td { padding: 7px 6px; } } </style> <div id="hunmin-model-card">

Hunmin-397B-A17B-CUA

A general-purpose 397B-parameter vision-language model (17B active) based on Qwen3.5-397B-A17B, developed through low-rank capability transfer followed by SFT and Agent RL (GRPO) for GUI grounding and multi-step desktop/browser tasks — substantially improving computer-use and fine-grained GUI-grounding performance while keeping Korean benchmark results close to the base model.

Technical Blog: Hunmin-CUA — Post-Training a 397B MoE for Computer Use

Overview

• Author: Hojun Lee (이호준) · LinkedIn
• Role: Project Lead & Primary Researcher
• Affiliation: GenON
• Base model: Qwen3.5-397B-A17B (MoE, 397B total / ~17B active)
• Post-training hardware: B200 x 8

Highlights

• Strong GUI grounding. Hunmin achieves 75.6 on ScreenSpot-Pro with single-view inference (76.6 with zoom-in inference), 96.2 on ScreenSpot-v2, and 86.8 on OSWorld-G-R, leading all models compared under the same reproduced evaluation setup.
• Large computer-use gains with Korean capability preserved. Compared with the base model, Hunmin improves by +22.3 on OSWorld and +9.1 on WindowsAgentArena. Across the Korean benchmarks reported below, performance remains within approximately 2 points of the base model.
• Practical computer-use setting. All computer-use evaluations use image_max = 5, matching the training/evaluation setup. We intentionally use a smaller visual history because image_max ≈ 20 can introduce substantial inference cost and memory overhead in real-world computer-use deployments.

Hunmin improves computer-use and GUI-grounding performance while keeping Korean benchmark results close to the base model.

Benchmarks

GUI Grounding

Author-reproduced results: baseline models were independently evaluated in the authors' environment rather than copied from model providers' reports.

Computer-Use Agents

All agent evaluations below use image_max = 5.

Author-reproduced results: baseline models were independently evaluated in the authors' environment rather than copied from model providers' reports.

Note: These agent scores should not be directly compared with results obtained using substantially different visual-history settings such as image_max ≈ 20.

Korean

Across the reported Korean benchmark suite, Hunmin remains close to the Qwen3.5-397B-A17B base model while substantially improving computer-use capability.

Gains over Base

The largest improvements appear on long-horizon agent tasks, while GUI grounding also improves and Korean performance remains close to the base model.

On OSWorld, the capability-transferred checkpoint achieved 66.2, and subsequent SFT and Agent RL (GRPO) further improved performance to 70.5 — an additional +4.3 points under the same evaluation setup.

Comparison with Qwen-CUA

Both models were evaluated under the same settings.

Developed through Qwen-CUA-derived low-rank capability transfer and subsequent post-training, Hunmin achieves a distinct capability profile. Under the same reproduced evaluation setup, it outperforms Qwen-CUA by 13.4 points on ScreenSpot-Pro and 3.0 points on OSWorld-G-R, while also achieving stronger results on several Korean benchmarks. Qwen-CUA remains stronger on long-horizon computer-use tasks.

For more detail about our evaluations, see our report.

Best Practices

• Coordinates: outputs are normalized to [0, 1000]; scale using /1000 × (width, height).
• Grounding: use native-resolution screenshots (up to ~16.7M pixels).
• Agents: use image_max = 5 to match the training and evaluation configuration.

Limitations

• Agent scores are configuration-dependent. Results reported here use image_max = 5 and should not be directly compared with evaluations using substantially different screenshot-history settings.
• Performance may vary with screenshot resolution, reasoning configuration, interaction limits, serving stack, and benchmark harness versions.

Quickstart (vLLM)

vllm serve <path>/Hunmin-397B-A17B-CUA \
  --tensor-parallel-size 8 \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.90 \
  --enable-prefix-caching

Training

Hunmin combines low-rank capability transfer from Qwen-CUA with subsequent SFT and Agent RL (GRPO). We computed the parameter delta between Qwen-CUA and the corresponding Qwen3.5 base checkpoint, approximated it using truncated SVD, and represented it as LoRA-style adapters. These adapters were first merged into Qwen3.5-397B-A17B BF16.

The resulting capability-transferred checkpoint was quantized to FP8 using the same fine-grained block-wise quantization scheme as the official Qwen3.5-397B-A17B-FP8 release. Further post-training was performed on this checkpoint using FP8 QLoRA.

The post-training adapters were merged into the BF16 counterpart of the capability-transferred checkpoint to produce the final model for release and evaluation.

• SFT data: OpenGVLab/ScaleCUA-Data
• RL environments: extreme1228/ScaleCUA · xlangai/CUA-Gym
• Training stack: MS-Swift · Ray · Megatron-Core · vLLM
• Hardware: 8× NVIDIA B200

Acknowledgements

Built on Qwen3.5-397B-A17B, with Qwen-CUA providing the source checkpoint for low-rank capability transfer and serving as a CUA-specialized reference model in our reproduced evaluations.

</div>