Web agents often gain task-specific skill by narrowing their focus; the harder trade-off is adding browser control without giving up the model’s general language and vision abilities. This release takes that broader approach, extending a Korean-English foundation model into visual web interaction.
Key Capabilities
- One model across language, vision, and action: It answers text queries, interprets images, and operates browser interfaces, avoiding a separate model for each step of a workflow.
- Screenshot-led interaction: The agent acts primarily from visual observations rather than requiring full DOM or accessibility-tree access, making it relevant to interfaces where those structures are unavailable.
- Bilingual web training: The developers report more than 10 million Korean-English web-pretraining samples and over 100,000 interaction trajectories. In their 612-task Online-Mind2Web evaluation, it scored 29.0% on English tasks and 30.1% on Korean tasks; these are the developers’ results, not an independent benchmark.
Who It’s For—and the Trade-off
A good fit for researchers and developers exploring Korean-English multimodal agents or browser automation that should retain general-purpose model behavior. Look elsewhere if web-task success alone is the priority: the developers say their specialized 4B and 9B variants perform better on web tasks, and the 11B model’s reported success rates leave substantial room for improvement.