Hunmin-CUA releases 397B open weights for computer-use capability transfer

GenON researchers have released Hunmin-397B-A17B-CUA, a vision-language mixture-of-experts model aimed at operating desktop and browser interfaces. The Hugging Face checkpoint is described as a 397-billion-parameter model with approximately 17 billion active parameters per token, derived from Qwen3.5-397B-A17B. It is listed with Safetensors artifacts and Apache-2.0 license metadata. The October 8 technical report addresses a practical problem in agent post-training: improving long-horizon computer-use behavior without sacrificing general visual grounding or Korean-language skills.

The researchers first tried supervised computer-use training and observed erosion in capabilities they wanted to retain. They instead extracted differences between a computer-use-specialized Qwen checkpoint and the related base, selected modules, approximated parts of the changes at low rank and merged them into the base before further supervised and reinforcement-learning stages. The report details different truncation ranks for attention, shared experts and routed experts; the approach is a targeted capability transfer, not a claim that a small adapter contains a fixed percentage of the model's skills.

The resulting numbers should be read carefully. Under the authors' setup, Hunmin scored 70.45 on a 360-task OSWorld evaluation, compared with 48.16 for their base and 77.12 for the specialized source checkpoint. Their ScreenSpot-Pro comparison reports 75.61 for Hunmin and 62.20 for the source. These are author-run comparisons, not Open Model Weights reproductions; the report also explains that a separate 316-task OSWorld partition must not be mixed numerically with the 360-task result.

The authors say the project used eight NVIDIA B200 GPUs for the training work and made explicit tradeoffs between computer-use performance, GUI grounding and Korean benchmarks. For builders, the interest is both the released checkpoint and a documented method for transferring specialized agent ability without blindly merging every changed weight. Deployment still demands substantial hardware planning and independent tests of the specific interface tasks, latency and safety constraints; a published checkpoint alone does not verify production readiness.