What happened
Microsoft’s Mage team published weights, a model card, and an OSWorld evaluation reference for Mage-VL-Cowork on Hugging Face. The repository was created on October 30, 2026. Mage-VL-Cowork is a 4B vision-language policy for computer use: it receives screenshot-grounded context and can issue mouse and keyboard actions. Unlike a policy confined to graphical interaction, it can also use code and terminal execution during the same episode.
That combined action space is the central product claim. A computer-use agent often encounters work that can be completed through visible interface operations, but some parts can be handled more directly through a command or a script. Mage-VL-Cowork is presented as a policy that chooses across those modes rather than treating GUI manipulation as the only way to act on a desktop or web task.
The project reports a 66.0% score on OSWorld-Verified under a 50-step budget. That figure is the average of two runs, 65.33% and 66.75%, across 361 held-out tasks; Google Drive tasks were excluded. In the project’s comparison table, Mage-VL-Cowork is ahead of Muse Glimmer-30B at 65.9% and Qwen3.6-27B at 63.9%, while SCALECUA-Qwen3.5-9B is listed at 68.7%. These are results reported by the developers themselves.
The release also describes its training recipe. Training data came from Mage-CUA-Forge, which synthesized 293,806 desktop and web tasks spanning 6,915 skills. The team says online reinforcement learning improved performance by 10.16 points over hybrid supervised fine-tuning, while reducing model calls by 13.0% and output tokens by 25.4%. The public release is available at https://huggingface.co/microsoft/Mage-VL-Cowork.
What the archive adds
This is a new fact on a known story. A prior Field Note, Microsoft releases Mage-VL-Cowork, a 4B hybrid GUI-plus-code computer-use model scoring 66.0% on OSWorld-Verified, identifies the release as distinct from earlier computer-use model reports because it combines GUI and code actions at 4B parameters. It also extends the context of Copilot for Windows hybrid intelligence, where Microsoft positioned Windows as a controlled execution layer using local and cloud AI.
Trend signal
The release supports harness architecture outlives model choice. Its practical contribution is not merely a reported score from a particular 4B model; it is a policy boundary that spans perception, GUI actions, and terminal/code execution. For production systems, the surrounding harness must still decide what tools are available, what each action can authorize, and how successful completion is verified.
Connections to other events
The earlier Copilot for Windows Field Note described Windows as an execution layer for agents that can combine local context with local and cloud models. Mage-VL-Cowork supplies a smaller open computer-use policy whose action design is consistent with that direction: a model is useful not only for interpreting a screen, but for selecting among bounded ways to operate on the environment.
Why it matters
For teams building AI workers, the release offers a concrete alternative to treating computer use as a long sequence of pixels-to-clicks decisions. A hybrid agent can reserve GUI interaction for tasks where the visible application state matters and use terminal or code execution where it is more direct. That can reduce both interaction length and the places where a visual action can fail.
The training results make that design worth testing even if the benchmark ranking does not transfer to a team’s environment. Microsoft reports fewer model calls and fewer output tokens after online RL relative to hybrid SFT. If those effects hold on a real workflow, the impact is operational: lower run cost, fewer turns to inspect, and potentially clearer recovery paths. Builders should evaluate task completion, permissions, and auditability rather than adopting the reported benchmark score as a procurement decision.
Limits and uncertainty
The supplied evidence does not establish independent performance. The OSWorld-Verified result, comparison table, and training gains are all reported by the project developers. A 50-step evaluation budget, two-run average, the exclusion of Google Drive, and a held-out set of 361 tasks define the stated result; they do not show how the policy behaves on a company’s applications, credentials, permissions, or failure conditions. The material also does not specify deployment hardware, latency, licensing terms, production support, or whether the published policy safely handles sensitive desktop environments.
Skill to build now
Build a small evaluation set where each task has two permitted routes: a GUI route and a terminal or code route. For each run, record the chosen route, completion evidence, number of model calls, output tokens, and any privileged action. This turns the hybrid claim into a measurable harness decision and reveals whether code execution reduces work or simply creates a different class of risk.
Reader takeaway
Mage-VL-Cowork is an open 4B computer-use release whose main design lesson is hybrid action selection, not its headline benchmark alone. Teams experimenting with desktop agents should test when visual interaction is necessary, when code is safer or cheaper, and how their runtime constrains both paths.