Zenon open-sources Hunmin VLM 397B, a computer-operating AI built on Alibaba's Qwen3.5-397B
The AMW Read
Zenon's efficient distillation of GUI-operating capability onto Qwen3.5-397B with only eight GPUs updates the open-weight computer-use leaderboard and evidences the efficient-capability-addition-over-scale argument its CTO stated explicitly.
Zenon open-sources Hunmin VLM 397B, a computer-operating AI built on Alibaba's Qwen3.5-397B
Seoul-based generative AI company Zenon (제논) released Hunmin (훈민) VLM 397B as an open-source model that views a screen, locates buttons and input fields, and directly carries out tasks across operating systems and applications. It is built on Alibaba's 397-billion-parameter Qwen3.5-397B, with added training for locating on-screen targets and completing multi-step tasks. On ScreenSpot-Pro, a benchmark for finding on-screen UI targets, it scored 75.6, ranking second among 48 models on a Hugging Face leaderboard and the only Korean-made model on that list. On OSWorld, a 360-task Linux benchmark, it scored 70.5, a 22.3-point gain over the base model, plus a 9.1-point gain on Windows tasks. Zenon says the model held within 2 points of the base model across eight Korean-language benchmarks including KMMLU, versus up to a 6-point drop for a comparison model built the same way. The full run — distilling an existing computer-use model's capability, shrinking model size, then applying supervised and reinforcement learning — used just eight Nvidia B200 GPUs, applying a small-GPU training method Zenon detailed in a July technical report for a separate 743-billion-parameter model. It ships under Apache 2.0 with an FP8 half-size variant and published re-evaluation documentation.
The release argues that adding a specific capability efficiently, not scaling parameter count, is becoming the competitive axis for open-weight computer-use models — the point Zenon's CTO made directly. That a Korean lab can take a large Chinese open-weight base model and bolt on GUI-operating ability with a single-digit GPU count, while holding local-language quality, is a concrete data point in the broader debate over whether frontier-scale compute is required to compete on agentic capability, and it puts a domestic player on a global GUI-agent leaderboard alongside far larger labs.
For builders, the FP8 release and documented re-evaluation method lower the bar for teams wanting a screen-operating agent without training one from scratch, particularly for Korean-language enterprise workflows where Zenon plans to route the model into its OneAgent (원에이전트) execution product. For investors tracking capital efficiency in model training, the eight-GPU claim for a near-400-billion-parameter capability upgrade is worth independent verification before treating it as a repeatable benchmark.
