A small AI model takes on real website tasks
KAIST researchers have released KORMo-Web, an artificial intelligence agent that operates websites through mouse clicks and keyboard input, with its smallest model completing 42.3% of assigned tasks in both Korean and English during an evaluation on real websites.
Contents
The team, led by Professor Lim Kyung-taek at KAIST's Graduate School of Culture Technology, announced the release on November 17, 2026. Its training effort used more than 10 million web screen samples and more than 100,000 records of website interactions to teach the system what it sees and how to act.
KORMo-Web can use services such as ChatGPT and Claude, search for products on Naver Shopping and Coupang, draft emails and carry out activities in online communities. The researchers also presented a demonstration video condensing 40 tasks on actual websites into four minutes.
The central idea is straightforward: instead of building a separate software connection for every service, give an AI the ability to understand the same screen a person sees and operate the controls on it.
KAIST released three model sizes, including a version with 4 billion parameters that can run directly on a personal computer. The release also includes model weights, training data, Korean evaluation data and code for a Chrome extension, giving developers the components needed to run and study the system.
How does an AI use a website like a person?
KORMo-Web is an AI agent, meaning it takes actions to carry out a request rather than simply returning a written answer. A user gives it an instruction, and the system interprets the web screen, selects controls and enters text to work toward the requested result.
A shopping task, for example, involves understanding where a search box is, entering a product name and navigating the results. Preparing material through another AI service requires finding its input area, submitting a request and handling the response.
Many software services provide an application programming interface, or API, through which another program can request information or perform an action. KORMo-Web instead aims to work through the visible website, reducing the need to build a dedicated connection for each service.
Lim described that goal when explaining the project:
Our goal is not to connect external application programming interfaces (APIs) one by one for each service, but to have AI understand the screen like a person and use the computer directly,
The approach also differs from automation built primarily around a page's underlying HTML or Document Object Model, known as the DOM. Those structures describe how a web page is organised. KORMo-Web's training centres on the displayed screen and records of how people use websites.
What the team taught it before asking it to act
The researchers built more than 10 million training samples from Korean and English web screens, including websites commonly used in South Korea. The aim was to teach the model more than the location of a button or text field.
The team calls this stage atomic level web pretraining. In simple terms, it means learning the small pieces of a website before trying to complete a larger job: where a menu is, what a button does, what an input field accepts and what state a control is in.
A button's appearance and its function are related but different information. An agent needs to identify the right control and understand how using it advances the user's task. Training on those details gives it a foundation for choosing an action.
The researchers then added more than 100,000 Korean and English web interaction records. These action sequences, also called trajectories, show how website use unfolds across several steps rather than in a single click.
The two training stages address different needs. Screen samples teach the agent to recognise and interpret what is in front of it. Interaction records teach it how a series of actions can lead from an instruction to a completed task.
What does a 42.3% success rate mean?
KORMo-Web-4B, the version with 4 billion parameters, achieved a task success rate of 42.3% in English and 42.3% in Korean on Online-Mind2Web, an evaluation that asks agents to perform tasks on real websites.
The official KORMo-Web-4B release page sets out the evaluation protocol and full results. All models ran the same 612 live-website tasks: 300 in English and 312 in Korean. Each model used its own prompts and action formats, while the same language model judge, identified as GPT-6, scored all runs. Success rates use the full task count as the denominator, so crashed or empty runs count as failures. The page also points to the full harness specification, system prompt and tool schema.
The full results, given as English and Korean success rates respectively, were Qwen3.5-4B at 20.0% and 13.5%; Fara-1.5-4B at 20.7% and 7.7%; UI-TARS-1.5-7B at 22.3% and 20.2%; OpenWebRL-4B at 32.7% and 15.4%; KORMo-Web-4B at 42.3% and 42.3%; and KORMo-Web-9B at 51.0% and 48.1%.
The researchers said this was the highest result among the agents with publicly available model weights included in their comparison. The model exceeded the previous best comparison result by 9.6 percentage points in English and 22.1 percentage points in Korean.
Those margins put the previous best results at 32.7% for English and 20.2% for Korean. The improvement was therefore much larger in Korean, where KORMo-Web's score was more than twice the previous comparison score.
A percentage point measures the gap between two percentages. Moving from 20.2% to 42.3%, for example, is an increase of 22.1 percentage points. That helps explain why the Korean result represents a substantial change within this particular evaluation.
The success rate also gives users a practical measure of reliability: at 42.3%, fewer than half of the assigned tasks were completed successfully. The remaining 57.7% fell outside the successful category, leaving considerable room to improve before users can routinely hand over work without checking the outcome.
Why the Korean evaluation data matters
KAIST also released Korean evaluation data adapted from four major benchmarks for graphical user interface agents. A graphical user interface is the collection of visual controls, such as buttons, menus and text boxes, through which people operate software.
People manually built the Korean evaluation data to suit Korean web environments. This gives researchers a way to test agents on local websites and language use, alongside the English settings that have dominated development.
The equal 42.3% scores on Online-Mind2Web are useful because they put Korean and English performance side by side under the reported evaluation. The larger gain over the previous Korean comparison result also makes the local language component a central part of the release.
Publishing the evaluation data allows other developers to test their systems against the same kinds of tasks. Combined with the training data and interaction records, it provides a clearer basis for investigating which parts of the training process improve performance.
Three models and the tools to run them
The release includes models with 4 billion, 9 billion and 11 billion parameters. Parameters are the internal numerical values a model learns during training; their number is a common way to describe model size.
The 4B and 9B versions are based on Qwen3.5. The 11B version, called KORMo-Web-11B, uses KAIST's own KORMo foundation model family as its starting point. A foundation model is a broadly trained AI system that can receive further training for a more specific role, in this case operating websites.
KORMo-Web-4B is intended to run directly on a personal computer. Its smaller size is central to the team's goal of making a web assistant available on a user's own machine rather than relying exclusively on a larger remote system.
The release goes beyond the models themselves. KAIST is making available the learned model weights, large collections of web training data, user action sequences, Korean evaluation data, Chrome extension code, installation guidance, videos and examples.
The researchers have not specified the minimum computer hardware requirements. The official release page does not state whether user confirmation is required before purchases or sending emails or public communications, such as comments. Those details matter to anyone considering using an agent for tasks that spend money or communicate with other people.
From demonstrations to useful delegation
The demonstrations cover several familiar kinds of work: preparing presentation material through ChatGPT, using Claude to investigate recent research, organising results in Notion and searching shopping websites. Email drafting and online community activity extend the system into tasks where the output may reach other people.
The four minute video presents 40 website tasks in condensed form. For prospective users, the more useful reliability measure is the separately reported task success rate, alongside the practical question of how much checking a particular job requires.
A draft email can be reviewed before sending. A product search can be checked against the user's budget and requirements. Publishing a comment or completing a purchase calls for a clear decision about when the user must approve the action. These differences make the choice of task as important as the ability to click the right controls.
Lim said his ambition is to bring this kind of assistance onto personal computers:
Going forward, we hope to create an environment where a small AI installed on a personal computer can perform a wide variety of web tasks on behalf of the user.
KORMo-Web brings that ambition into a form others can examine and run: three released models, extensive training resources and Korean testing tools. Its current results show progress in completing real website tasks, with successful execution still below half of the tasks in the reported evaluation.
Key Points
- KAIST announced KORMo-Web on November 17, 2026, as an AI agent that operates websites through visible controls.
- Training used more than 10 million Korean and English web screen samples and more than 100,000 interaction records.
- The 4B model achieved 42.3% task success in both languages on Online-Mind2Web.
- Its lead over the previous best comparison model was 9.6 percentage points in English and 22.1 in Korean.
- The release includes three model sizes, Korean evaluation data, training resources and Chrome extension code.
- The smallest model can run on a personal computer, while its reported success rate leaves substantial work for users to check.






