Microsoft’s open 4B agent clicks and codes, and scores 66% on OSWorld

Microsoft has published Mage-VL-Cowork, an open 4-billion-parameter model that operates a computer through both screenshots and a code terminal. Microsoft says its 66.0% on OSWorld-Verified is the best result among 4B models.

One checkpoint that routes between mouse and shell

The model card describes a single vision-language policy with two action channels. computer_use covers 11 GUI primitives, such as clicking and typing, in normalized coordinates. bash gives a persistent shell and Python environment. Microsoft says the model picks a channel on each turn without an external router. It reaches for the GUI for visual layouts and canvas editing, and for the shell for structured edits to spreadsheets, documents and presentations.

It is built on Microsoft’s Mage-VL 4B base, combining the Mage-ViT vision encoder with Qwen3-4B-Instruct-2507. The weights total 9.48 GB in bfloat16. The active context window is 32,768 tokens, though the card says the architecture supports up to 262,144. To save context, the model keeps the three most recent screenshots and four full responses, and folds older turns into one-line action summaries.

Training ran in five stages: GUI midtraining, GUI agent fine-tuning, coding and terminal fine-tuning, a hybrid fine-tuning stage, and online reinforcement learning. Microsoft lists 32 Nvidia B200 GPUs plus 1,120 CPU sandbox slots for the reinforcement learning. According to the card, the reinforcement learning stage lifted OSWorld-Verified accuracy from 55.84% to 66.04% while cutting mean output tokens per task from 7,374 to 5,499.

That figure comes from the teacher setup. Microsoft pruned overlapping tools such as xdotool and accessibility-tree dumps from the GPT-5.4 teacher’s harness and pre-installed Python libraries such as openpyxl and pandas. It reports the teacher’s success rose to 92.0% with 29.2% fewer steps.

Best at 4B, but not the top of the table

The 66.0% is a two-run average (65.33% and 66.75%) on 361 OSWorld-Verified tasks, excluding Google Drive, with a 50-step budget. Microsoft’s own comparison table shows where that sits. The headline claim is limited to 4B models: the only other 4B entry, Fara1.5-4B, scores 13.9%.

Larger and proprietary systems score higher. ScaleCUA-Qwen3.5-9B is listed at 68.7% and GPT-5.4-mini at 69.6%. The card also lists GPT-5.5 at 78.7%, Claude Opus 4.7 at 78.0% and a human reference of 72.4%. Mage-VL-Cowork does lead on the operating-system category (87.5%) and scores 62.7% on multi-application workflows. It trails ScaleCUA on office, daily and professional categories.

Everything here is Microsoft’s own reporting. The page does not point to independent reproduction. It does ship an osworld_eval/ harness and SGLang serving instructions so others can check the number. It also prints two slightly different gains from the first stage to the last, +20.44 and +20.48 points, which looks like rounding between tables. The release also carries a Responsible AI section, and Microsoft’s tags list the model for GUI, coding and terminal agents.

Open weights under Apache 2.0

The checkpoint is public and ungated, licensed Apache-2.0, and loads through Transformers with custom code. The page showed zero downloads at the time of publication, so no community results exist yet. Microsoft does not state hardware requirements for running it. The only compute figures given are for training. The card links the base model microsoft/Mage-VL and the vision encoder microsoft/Mage-ViT as part of the same family.

The savings are in tokens and in data, not in headline score

The economics angle is the more interesting one. Microsoft compares Mage-VL-Cowork with GPT-5.4-mini at 18.77 versus 18.24 model calls per task, but with 5,499 output tokens against 12,397. That is about 55.6% fewer, and it matters for any team paying per token to run a desktop agent at scale. Because the model is small enough to self-host, there is no per-call fee at all, though Microsoft publishes no serving cost.

The data side shows the same thinking. Microsoft says its Mage-CUA-Forge pipeline builds reusable generation workflows once per skill instead of running an open-ended agent loop for each task. That cut tokens per task from 1,276K to 85K and recurring cost from $0.906 to $0.158. The result is 293,806 synthesized tasks and 318,837 verified trajectories. Those are the company’s figures, produced with a model it calls gpt-5.6-terra.

The release also fits Microsoft’s broader spending picture. Our coverage of its quarter reported that Azure grew 48% against a $677 billion backlog. We also reported on a $25 billion bond plan to fund the AI buildout. Against that scale, a 4B agent trained on 32 GPUs shows Microsoft also working on the opposite end, where an agent is cheap enough to run locally.

What the page does not establish is how the model behaves outside its benchmarks. WebVoyager-Static is the only browser test the visible text mentions, and the result is cut off. There is also no information on failure modes or on safeguards for an agent that can run shell commands.