Mage-VL-Cowork: 4B Agent Learns GUI and Code Use in One Policy

A research team has presented Mage-VL-Cowork, a 4-billion-parameter agent designed to handle graphical interfaces and code execution as part of the same decision-making policy. The work matters because it treats computer use as a combined skill, rather than forcing a choice between clicking through an application and writing commands in a terminal.

According to arXiv, the paper was submitted on 29 October 2026 by Xiaoyi Zhang and 21 other authors. The authors describe the model as a hybrid computer-use agent, or CUA, built for tasks that span desktop and web environments. Their central claim is that a relatively small model can become more capable on practical computer tasks when training, data generation, and teacher supervision are designed together, instead of being assembled as separate pieces.

What is Mage-VL-Cowork?

Mage-VL-Cowork-4B is presented as a single policy that can move between visual interaction and programmatic control. In a typical computer-use setting, an agent may need to read what is on screen, decide which control to select, open a file, change a setting, or run a command to complete a job more reliably than clicking would allow. The authors argue that effective agents should be able to combine those modes without handing off to a separate system.

That framing is important for how the model was built. Rather than starting from a model specialised only for language or only for vision, the team started from a checkpoint that was already aligned to multimodal input. From there, they applied a progressive training programme intended to teach computer use step by step. The paper positions this as an end-to-end recipe: the starting model, the curriculum, the generated tasks, and the supervision used at each stage are all part of the same design.

The authors also describe the result as a step for on-device intelligence. That claim rests on the model's size. At 4B parameters, Mage-VL-Cowork is far smaller than many general-purpose frontier systems, so the practical question is not only whether it can complete tasks, but whether similar capability can be delivered with lower reasoning cost and in settings where a smaller model is easier to run.

How was the model trained?

According to arXiv, training followed a five-stage curriculum spread across three broad phases: midtraining, supervised fine-tuning, and reinforcement learning. The authors call this a progressive curriculum, meaning the model was not expected to master full computer-use behaviour in one pass. Instead, capabilities were introduced and refined across successive stages, with supervision matched to what each stage was meant to teach.

That staged approach is one of the paper's main contributions. Many agent projects rely on a single round of fine-tuning on demonstrations, followed by evaluation. Here, the sequence is presented as deliberate: an initially general multimodal model is gradually shaped into a computer-use policy, with later stages building on earlier ones. The authors say the model climbed to its final benchmark result through these different training stages, suggesting that the gains were cumulative rather than coming from one final adjustment.

The supervision strategy also differed from straightforward imitation. The authors state that they did not simply distil a default teacher. Instead, they evolved the teacher harness — the surrounding setup used to prompt, constrain, and collect behaviour from the teacher — and then created specialised harnesses for computer use. Those harnesses produced trajectories, or recorded sequences of actions and outcomes, alongside supervision tailored to each stage.

For practitioners, the lesson is methodological as much as it is about the final model. If the harness that generates training behaviour is weak, the student inherits its habits and its blind spots. By improving the harness first and then specialising it for particular kinds of computer use, the team aimed to raise the quality of the data before the student model ever learned from it.

What is Mage-CUA-Forge?

To supply training at scale, the authors built Mage-CUA-Forge, a system for generating computer-use tasks. The problem it addresses is familiar to anyone who has tried to create agent training data: building a fresh agent setup for every individual task is slow and expensive, and the resulting data can be inconsistent. Mage-CUA-Forge replaces that per-task construction with reusable generation workflows conditioned on skills, together with preparers tailored to specific applications.

According to arXiv, that approach produced 293,806 desktop and web tasks covering 6,915 skills. The scale is central to the paper's argument. A curriculum that moves through several stages needs varied examples, and a smaller set of hand-written tasks would be unlikely to cover the range of applications and abilities the authors wanted to teach. Skill-conditioned generation is intended to make that coverage systematic, while application-specific preparers handle the setup details for particular software.

The authors also report a cost saving: recurring per-task expenditure fell by 82.6%. That figure matters because data generation is often the hidden cost of agent training. If each new task requires bespoke engineering, scaling the dataset becomes difficult even when the underlying model is small. A reusable pipeline that lowers the marginal cost per task makes larger and more diverse datasets more practical, and it makes the training recipe easier to repeat or extend.

What results did it achieve?

The headline result is 66.0% on OSWorld-Verified, a benchmark for computer-use tasks in operating-system environments. According to arXiv, Mage-VL-Cowork-4B reached that score after progressing through its training stages, and the authors describe it as state-of-the-art performance at its parameter size. They also say the model achieved this with more efficient reasoning cost, which they link to its suitability for on-device use.

Two qualifications are worth keeping in mind. First, the state-of-the-art claim is scoped to the model's parameter size, not to computer-use agents in general. A 4B model leading its size class is a different statement from leading the entire field, and the paper's framing is careful to make that distinction. Second, OSWorld-Verified is one benchmark. Strong performance there is evidence of computer-use capability, but it does not by itself show how the model would handle every application, interface, or workflow a real user might present.

Even with those limits, the result is notable for the combination it represents: a small model, a multi-stage curriculum, a large generated task set, and a benchmark score that the authors present as leading for its size. The paper's argument is that these elements reinforce each other. Better harnesses produce better trajectories, better trajectories support a more effective curriculum, and the curriculum allows a compact model to reach performance that might otherwise require a larger system.

How does this relate to Grafy?

The closest Grafy connection is Operator, Grafy's browsing agent for web tasks, because both approaches concern agents that act in an interface rather than only answering questions in text. Mage-VL-Cowork is aimed at desktop and web computer use, while Operator is a browsing agent, so the overlap is in the broader idea of an agent completing tasks through an environment. Readers interested in how such agents are used in practice can compare the paper's emphasis on combined visual and code-based action with Operator's web-task focus.

There is also a genuine connection for developers through Grafy Code, an AI coding studio. The paper's hybrid framing — combining interface interaction with terminal and code execution in one policy — speaks directly to workflows where an agent writes code, runs it, inspects the result, and adjusts its approach. Grafy Code is not the model described in the paper, but it is a place where readers can explore the code-execution side of hybrid agent work.

For researchers who want to study the paper itself, Reader offers a way to read, search, and question PDFs and papers with AI. That makes it a practical starting point for checking the curriculum design, the Mage-CUA-Forge pipeline, and the OSWorld-Verified evaluation in the original text, rather than relying on a summary alone.

Frequently asked questions

How many tasks did Mage-CUA-Forge synthesize?

According to arXiv, Mage-CUA-Forge synthesized 293,806 desktop and web tasks. Those tasks covered 6,915 skills. The system used reusable, skill-conditioned generation workflows and application-specific preparers, rather than constructing a separate agent setup for each task.

What score did Mage-VL-Cowork-4B achieve on OSWorld-Verified?

Mage-VL-Cowork-4B achieved 66.0% on OSWorld-Verified, according to arXiv. The authors describe this as state-of-the-art performance at its 4B parameter size. They also say the model reached the result through its different training stages with more efficient reasoning cost.

How many stages were in the Mage-VL-Cowork training curriculum?

The training curriculum had five stages. According to arXiv, those stages were progressive and spanned midtraining, supervised fine-tuning, and reinforcement learning. The model was trained from a general multimodally aligned checkpoint through that end-to-end sequence.

Source: arXiv Computation and Language, arXiv:2610.40747 — https://arxiv.org/abs/2610.40747

---

Source: arXiv Computation and Language — Mage-VL-Cowork: A Full Training Recipe to Learn Hybrid Computer Use From Evolved Harnesses