The spreadsheet needs two merged headings and a chart. One agent starts clicking through menus. The other writes a short program that edits the workbook. In Microsoft's side-by-side demonstration, the second route finishes in 39 seconds and six steps. The clicking route takes nearly ten minutes.

That is a selected demo, not a promise about your next spreadsheet. But it makes the interesting part of Mage-VL-Cowork easy to see: a computer-use model can learn when the screen is the right interface, and when it is an expensive detour.

A smaller model with two ways to work

Microsoft's Mage team has released Mage-VL-Cowork, described in its model card as a 4-billion-parameter vision-language model. One policy chooses between screenshot-based mouse and keyboard actions and a persistent terminal with Python. It does not need a separate dispatcher to make that choice.

For a workbook, code can change cells directly. For a visual edit, the interface may still be the better route. Microsoft's second demonstration uses GIMP, then checks the saved file through the terminal. Neither example establishes a universal speed advantage, and the comparison gives the two agents different action spaces: hybrid for Mage, GUI-only for its rival.

The weights are available under Apache-2.0. The model card lists 9.48 GB of bfloat16 weights and an active 32,768-token window. A small download is not a ready-made office assistant. Running a checkpoint, providing a desktop, handling its actions and checking the resulting files remain separate jobs. OnPrem World's release coverage makes the local-running appeal clear; it should not be mistaken for proof that any ordinary laptop can run the full setup well.

Read the 66% with its conditions attached

Microsoft reports 66.0% on OSWorld-Verified within a 50-step budget. The model card's protocol specifies 361 tasks, excluding Google Drive, at a 1920-by-1080 desktop resolution. The headline is the average of two full runs, reported as 65.33% and 66.75%.

Those are the team's results, not GigSoul measurements. They do not mean two-thirds of an employee's work can now be handed over. Nor is this the best score in the team's own comparison table: the larger ScaleCUA-Qwen3.5-9B is listed at 68.7%. The useful claim is narrower: a compact model is competitive on this particular desktop test. i-SCOOP's analysis likewise separates the research checkpoint from a finished product.

The teacher's setup leaves fingerprints

The research offers another result worth watching. With the same teacher model and 672 held-out multi-application tasks, Microsoft compares GUI-plus-code against a version with an added accessibility channel. The teachers score 92.0% and 89.6%. The smaller students trained from their respective demonstrations score 37.6% and 25.1%.

That is evidence for this experiment, not a rule against accessibility tools. It suggests that extra action routes can complicate what a smaller model learns. The surrounding setup is part of the system being evaluated, not background furniture.

Keep the experiment disposable

The model card warns that this agent can execute arbitrary shell commands. Its guidance calls for isolated containers or virtual machines, no access to host files or personal environment variables, restricted networking, and explicit human confirmation for consequential actions. Credentials should not be casually mounted into the workspace.

For someone considering a trial, the first question is therefore less glamorous than the leaderboard: can the agent finish a copied task in a disposable environment, and can you check the saved result? GigSoul has not run that trial. The release is interesting because it gives researchers a compact system to investigate that question, not because a quick spreadsheet demo makes supervision obsolete.