Microsoft published Mage-VL-Cowork on Hugging Face on October 30, a 4-billion-parameter model that operates a computer by choosing, at each step, between clicking through the screen and running code or shell commands. The weights are free to download under the Apache-2.0 license, and Microsoft's Mage team reports that the model completes 66.0% of the tasks in OSWorld-Verified, a desktop-automation benchmark on which people manage 72.4%.
SaasBuzz found no Microsoft blog post or launch event for the release. The team posted the model card, the code on GitHub, a project page and a 30-page technical report. Every score in them is Microsoft's own, and SaasBuzz has not run the model.
How it works
The model reads a screenshot and a task, writes its reasoning, and then calls one of two tools. One, computer_use, clicks, drags, types and scrolls on the screen. The other, bash, runs commands and Python scripts in a shell. The model card says the model picks the screen for "visual layouts, canvas editing, and live in-memory application state" and the shell for "structured programmatic edits (spreadsheets, documents, presentations)", with no separate router deciding for it.
In the card's example, asked to merge cells in a spreadsheet and save it, the model skips opening LibreOffice and edits the file with a Python library from the shell. Microsoft says that on one such task it finished in 39 seconds and six steps, while Qwen3.6-27B, a 27B model working only through the screen, took nine minutes and 56 seconds. That is one demonstration chosen by Microsoft, not a benchmark.
Underneath is Microsoft's own Mage-ViT image encoder joined to Qwen3-4B, a small language model. According to the technical report, the team trained it on 318,837 recorded runs by a "teacher" agent across 293,806 synthesized tasks.
How the 66% was measured
The team ran the model on 361 of OSWorld's 369 Linux desktop tasks, leaving out eight that need Google Drive, which the benchmark's maintainers allow. Each task had a 50-step limit at 1920 by 1080 resolution. The 66.0% is the average of two full runs, which scored 65.33% and 66.75%.
Microsoft's comparison table puts the model ahead of much larger open models, though not all of them.
Model | Size | Score |
|---|---|---|
GPT-5.5 | Not given | 78.7% |
ScaleCUA-Qwen3.5-9B | 9B | 68.7% |
Mage-VL-Cowork | 4B | 66.0% |
Muse Glimmer-30B | 30B | 65.9% |
Qwen3.6-27B | 27B | 63.9% |
Source: Microsoft's Mage-VL-Cowork model card, OSWorld-Verified overall scores. All figures are as Microsoft lists them.
Two things in that table deserve care. The model's lead over Muse Glimmer-30B is 0.1 points, while its own two runs differed by 1.42 points, so the two are effectively level. And the comparison is selective: the OSWorld-Verified leaderboard, where the benchmark's maintainers post results they have checked, lists 22 entries at or above 66.0% as of October 31, and every one of them was allowed 100 steps rather than 50. Mage-VL-Cowork is not on it yet. Among the leaderboard's single-attempt entries limited to 50 steps, the best is 65.77%, so Microsoft's figure would lead that group if the maintainers confirmed it.
What builders need to run it
The weights take 9.48 GB in bfloat16, and the card's serving example runs on a single GPU, though it names no minimum card. Microsoft recommends serving the model with SGLang and points to a prebuilt Docker image on a Docker Hub account named xiaoyizhang, which matches the name of the technical report's first author, Xiaoyi Zhang, rather than on a Microsoft account. The model uses custom code, so loading it requires trust_remote_code=True.
The card is blunt about where the model belongs. Because it can run "arbitrary shell commands", Microsoft says it is "designed and evaluated for isolated, disposable sandbox environments", and asks for human confirmation before irreversible steps such as payments or sign-ins.
What it means for teams building desktop agents
Of the 22 leaderboard entries at or above Microsoft's score, 12 wrap a model in an agent framework, and all used twice the step budget. Microsoft's argument is that a 4B model that knows when to drop to the shell can get close on its own: the card reports 18.77 model calls per task, against 18.24 for GPT-5.4-mini, while producing 55.6% fewer output tokens. If outside tests hold that up, an automation team could keep such an agent on its own hardware instead of sending every screenshot to a hosted model. SaasBuzz has covered OpenAI's and Anthropic's research on monitoring AI agents. With a downloadable agent, that monitoring falls to whoever runs it.
What is not yet known
SaasBuzz found no independent reproduction by midday ET on October 31. The technical report sits on GitHub, and an arXiv search that day found no listing, though the card's badge for it is labeled arXiv. The tests cover Linux desktops, and the card lists English only, so performance on Windows or macOS is not measured. And Microsoft has not said whether Mage-VL-Cowork will reach any of its products.
