Most computer-use agents work like a person with a mouse: look at a screenshot, click, repeat. Microsoft’s new Mage-VL-Cowork takes a shortcut when it can. The 4-billion-parameter open model can switch between clicking through an app and running code in a terminal, and Microsoft reports 66.0% on the OSWorld-Verified desktop benchmark. The weights went up on Hugging Face on October 30, 2026, under the Apache-2.0 license.

Coffee summary

  • FACT: Mage-VL-Cowork is a 4B vision-language model from Microsoft’s Mage team, released under Apache-2.0.
  • FACT: It has two action channels: computer_use for GUI clicks and typing, and bash for shell and Python commands.
  • CLAIM: Microsoft reports 66.0% on OSWorld-Verified (361 tasks, excluding Google Drive, 50-step budget, average of two runs), which it calls the best result at the 4B scale.
  • CLAIM: After reinforcement learning, Microsoft says the model uses 13.0% fewer model calls and 25.4% fewer output tokens than its earlier training stage.
  • FACT: Microsoft’s model card says to run it only in isolated sandboxes.

What happened

Microsoft published the model, a model card, and an evaluation package on Hugging Face, plus a technical report. The model is built on a Qwen3-4B language backbone with Microsoft’s own Mage-ViT image encoder. It ships as about 9.48 GB of bfloat16 weights and uses a 32,768-token working context.

The key idea is “hybrid” computer use. For visual tasks, like editing an image canvas, the model clicks. For structured edits, like changing a spreadsheet, it writes a short Python script instead of dragging through cells. It makes that choice itself, step by step, without a separate router.

Why it matters

OPINION: Clicking through apps one screenshot at a time is slow for AI agents. For many file tasks, code is faster and easier to verify. A small model that knows when to use each could make desktop agents cheaper to run.

CLAIM: Microsoft’s own demo shows the gap. In a LibreOffice Calc task, Mage-VL-Cowork finished in 39 seconds and 6 steps by editing the file with Python. Qwen3.6-27B, a GUI-only model, took 9 minutes 56 seconds and 39 steps. That is one vendor demo, not a broad benchmark.

How it compares (Microsoft’s numbers)

Model OSWorld-Verified (as listed by Microsoft)
Human reference 72.4%
GPT-5.4-mini (one run, same 361 tasks) 69.6%
ScaleCUA-Qwen3.5-9B (9B) 68.7%
Mage-VL-Cowork (4B) 66.0%
Qwen3.6-27B (27B) 63.9%
Fara1.5-4B (4B, browser specialist) 13.9%

The top proprietary models in the same table (GPT-5.5, Claude Opus 4.7, and GPT-5.4) score higher, from 75.0% to 78.7%. Not every row uses the same step budget or task set, so treat the table as a rough guide.

Who should care

  • Developers building desktop agents who want an open, commercially usable model.
  • Teams automating office files such as spreadsheets, documents, and slides.
  • Researchers interested in the training recipe: Microsoft describes a five-stage curriculum and a synthetic task engine that generated 293,806 tasks.

Limitations

  • Self-reported scores. The benchmark numbers come from Microsoft and have not been independently reproduced yet.
  • Not plug-and-play. It uses a custom architecture, so you must load it with trust_remote_code=True, and Microsoft publishes recommended sampling settings and a serving setup.
  • “On-device” is a goal, not a guarantee. Microsoft says the model pushes on-device intelligence, but 9.48 GB of weights still needs a capable GPU.
  • Real risk. A model that runs shell commands can delete files or leak data if misused or tricked by a malicious web page.

What to do next

  1. Read the model card on Hugging Face before testing, especially the Responsible AI section.
  2. Sandbox it. Microsoft says to use an isolated container or virtual machine, with no access to host files, local networks, personal environment variables, or credentials.
  3. Add human approval for irreversible actions such as payments, logins, or sending data out.
  4. Start with file tasks where results are easy to check, like spreadsheet edits, before trying open-ended web work.

AIImpish Take

Mage-VL-Cowork’s real contribution is the habit it teaches: use code when code is the better tool. If the results hold up independently, small open agents could take on a lot of routine office work at low cost. Builders should test it now, but only inside a locked-down sandbox.

Related on AIImpish

*Sources: Microsoft, Mage-VL-Cowork model card on Hugging Face; Microsoft Mage project page; Mage-VL-Cowork technical report (PDF).*