Microsoft Mage-VL-Cowork computer use is the strongest small open-weight desktop agent to appear this month. On October 30, 2026, Microsoft’s Mage team published the weights for Mage-VL-Cowork, a 4-billion-parameter vision model that drives a computer by choosing, step by step, between clicking the screen and running terminal code.
That choice is the whole idea. Most computer-use agents pick one channel: they click like a human, or they script like a developer. Mage-VL-Cowork does both in one model, and Microsoft reports 66.0% on OSWorld-Verified, a benchmark of real desktop tasks.
How the hybrid actually works
The model sees a screenshot and an English task, reasons about the next move, and answers with one action: a GUI action from its 11-action toolkit, or a bash command in a persistent shell.
Under the hood, a vision encoder called Mage-ViT reads the screen, wired to Qwen3-4B-Instruct-2507, Alibaba’s text model. The starting point is Mage-VL, uploaded by the Mage team in July 2026.
One project-page demo shows why the mix matters: the agent finished a task in 6 steps and 39 seconds, almost entirely with Python in the terminal, touching the interface only once. A click-only agent would have taken the slow road through menus.
The specs are modest: about 4 billion parameters, 9.48 GB of weights, and a 32,768-token active context window — small enough for a single capable GPU, not a data center.
The 66% number — read the fine print
Microsoft’s 66.0% is a two-run average on OSWorld-Verified (361 tasks, 50-step budget, Google Drive tasks excluded), scored by the team itself with its own agent at 1920×1080. The runs scored 65.33% and 66.75%.
That is the best self-reported result among 4B models in Microsoft’s comparison, beating larger open models in the same table, including a 30B model at 65.9% and a 27B model at 63.9%. But the caveats matter:
- Self-reported. Nobody independent has repeated the score yet, and the model is not on the official OSWorld leaderboard.
- Not the overall best. A 9B model scored 68.7% in the same comparison, and the official leaderboard’s best reaches 86% with twice the step budget.
- Linux only. The evidence is a Linux desktop. Windows, Outlook, SAP and line-of-business software were not tested.
- Not the best web agent. An earlier checkpoint scored 58.6% on an offline web benchmark, behind a browser-only specialist at 68.0%.
The licence is Apache-2.0, which permits commercial use — but Microsoft states the Mage models are for research purposes only, not for product or service deployment. That sentence is what a business should read before building on it.
The evidence: model facts
| Detail | What the source says |
|---|---|
| Published | October 30, 2026, on Hugging Face, by Microsoft’s Mage team |
| Size | About 4 billion parameters; 9.48 GB of weights |
| Architecture | Mage-ViT vision encoder plus Qwen3-4B-Instruct-2507 language backbone |
| Actions | 11 GUI primitives plus a persistent bash and Python shell, one policy |
| Headline score | 66.0% on OSWorld-Verified, 50 steps, two-run average, 361 tasks |
| Benchmark caveats | Self-reported; Drive tasks excluded; not on leaderboard |
| Larger-model comparison | Ahead of a 30B model at 65.9% and a 27B model at 63.9% in Microsoft’s table |
| Higher scores elsewhere | 9B ScaleCUA 68.7%; leaderboard best 86% (100 steps) |
| Licence | Apache-2.0, with a research-purposes-only statement from Microsoft |
OSWorld-Verified is a benchmark that asks an agent to finish real tasks inside a real operating system — spreadsheets, image editors, browsers, files — and checks the result with a script.
Should you build on Mage-VL-Cowork? A 4-step check
1. Is your task Linux desktop work? — That is what was tested. Windows business software was not.
2. Does it mix clicks and code? — The hybrid wins where some steps are faster as commands than as clicks.
3. Can you sandbox it? — It runs shell commands. Isolated containers or virtual machines are mandatory, not optional.
4. Does the research-only note block you? — Read Microsoft’s statement before promising a client a product built on it.
How to profit from it
A small, free, capable desktop agent is a service opportunity precisely because it is not finished. Clients pay for someone to make it safe and useful.
- Sell sandboxed desktop automation. Offer Linux desktop workflows — file sorting, spreadsheet prep, report assembly — run inside isolated containers, with a human approving anything irreversible.
- Charge for the evaluation, not the model. Run a client’s actual tasks against the model and report the real success rate. That honest number is worth more than Microsoft’s 66%.
- Build hybrid workflows. The click-plus-code mix suits jobs that are half GUI, half script: data entry from a legacy app into a spreadsheet, for example.
- Never run it on a live machine. An agent with shell access outside a sandbox is a liability. Put isolation in the contract.
- Watch for the independent score. When an independent lab repeats the benchmark, update your pitch the same day.
Sources
Microsoft’s Mage team model card and technical report as covered by ExplainX and I-Scoop. Check the Hugging Face model card itself before downloading or deploying anything.
Frequently Asked Questions
What is Mage-VL-Cowork?
A 4-billion-parameter open-weight vision-language agent from Microsoft’s Mage team that operates a computer through GUI clicks and terminal commands.
How does it decide between clicking and coding?
A single model policy chooses, turn by turn, whether the next action is a GUI action or a bash command.
What does 66.0% on OSWorld-Verified mean?
In Microsoft’s own two-run test, the agent completed 66.0% of 361 verified desktop tasks within 50 steps each.
Is the score independently verified?
Not yet. The numbers are Microsoft’s own, and the model is not on the official OSWorld leaderboard.
Can I use it commercially?
The weights are Apache-2.0, but Microsoft states the Mage models are for research purposes only and not intended for product or service deployment. Read that statement first.
Keep Reading
Want more tools that actually pay for themselves? Browse the AI Tools section for the latest launches, pricing breakdowns, and profit playbooks — and check the official source linked above before you buy or deploy anything.