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.

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

Benchmark	Base	Hunmin (ours)	Qwen-CUA
ScreenSpot-Pro	72.7	75.6	62.2
ScreenSpot-Pro (zoom-in)	75.1	76.6	71.9
ScreenSpot-v2	95.2	96.2	95.0
OSWorld-G	73.2	79.9	75.7
OSWorld-G-R	79.2	86.8	83.8

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.

Benchmark	Base	Hunmin (ours)	Qwen-CUA
OSWorld (360)	48.2	70.5	77.1
WindowsAgentArena (153)	41.8	50.9	57.3

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

Benchmark	Base	Hunmin (ours)	Qwen-CUA
KMMLU-Redux	82.6	81.4	78.1
KMMLU-Pro	77.9	76.1	71.4
IFEval-Ko	84.7	85.1	81.0
K-MMBench	92.5	92.5	90.3
K-SEED	81.3	81.5	80.9
K-MMStar	81.1	80.3	75.0
K-DTCBench	98.3	98.2	97.5
K-LLaVA-W (judge)	126.0	128.0	130.1

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

Benchmark	Base	Ours	Δ
OSWorld (360)	48.2	70.5	+22.3
WindowsAgentArena (153)	41.8	50.9	+9.1
OSWorld-G-R	79.2	86.8	+7.6
ScreenSpot-Pro (single)	72.7	75.6	+2.9

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.

Benchmark	Qwen-CUA	Ours	Δ (Ours - Qwen-CUA)
ScreenSpot-Pro	62.2	75.6	+13.4
OSWorld-G-R	83.8	86.8	+3.0
KMMLU-Pro	71.4	76.1	+4.7
K-MMStar	75.0	80.3	+5.3
OSWorld (360)	77.1	70.5	-6.6

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.