Researchers detail Mage-VL-Cowork, a 4B hybrid computer-use AI agent

Researchers have published a full training recipe for Mage-VL-Cowork, a 4-billion-parameter computer-use agent that combines screen clicks, written code and terminal commands inside one policy. The paper, posted to arXiv by a 22-author team that includes Xiaoyi Zhang and Haoqing Wang, reports that the model reached 66.0% on OSWorld-Verified, a benchmark used to evaluate computer-use agents. The authors say that score is state-of-the-art for a model of its size.

Computer-use agents are AI systems built to operate a computer the way a person does: reading a screen, clicking buttons, typing commands and running scripts. The paper frames most existing systems as leaning on a single mode of action, either screen control or code execution, rather than combining both inside one policy. Mage-VL-Cowork is designed to switch between the two, and the authors trained it through five stages, moving from midtraining, an additional pretraining stage, through supervised fine-tuning, called SFT, to reinforcement learning, called RL. Each stage starts from what the paper describes as a general multimodally aligned checkpoint, a model already trained to connect text and images.

To supply enough training examples, the team built a system called Mage-CUA-Forge. Instead of building each task by hand, it uses what the authors call skill-conditioned generation workflows, templates tied to a named skill, together with application-specific preparers that set up each task’s environment. The paper says this pipeline synthesized 293,806 desktop and web tasks across 6,915 skills, while cutting the recurring cost of producing each task by 82.6%.

Evolving the teacher

Rather than copying a single fixed teacher model’s outputs, the researchers say they evolved the teacher harness itself, the surrounding code that lets a model take actions and receive results, into specialized versions built for different computer-use jobs. Those specialized harnesses produced the step-by-step action records, called trajectories, used to supervise Mage-VL-Cowork at each stage of training, according to the paper.

The authors present the 66.0% OSWorld-Verified score as evidence that a 4-billion-parameter model, small enough to plausibly run on a single device, can reach state-of-the-art performance for its size while using what they describe as more efficient reasoning cost. They frame this as progress toward on-device intelligence, AI capable of running locally rather than through a remote server.

The paper does not name the authors’ employer or funding source, and it has not yet completed peer review; its arXiv-issued DOI is listed as pending registration. It is also not yet clear whether Mage-VL-Cowork’s model weights, the Mage-CUA-Forge pipeline, or the training data will be released publicly, or whether independent groups will be able to reproduce the OSWorld-Verified score. Watch for a public code or model release, and for outside attempts to replicate the benchmark result.