Mage-VL-Cowork
A 4B computer-use agent from Microsoft that decides when to click and when to run code. 66.0% on OSWorld-Verified, Apache 2.0.
In short. Mage-VL-Cowork looks at a screenshot and operates a computer: it clicks, types, and runs shell commands and Python. Because it is only 4 billion parameters, it can run on a single GPU. It is free for commercial use under Apache 2.0, but it takes real actions, so run it in a sandbox.
Summary
Mage-VL-Cowork is a 4-billion-parameter model from Microsoft's Mage team that operates a computer. It reads screenshots and, within one conversation, chooses between two tools: graphical actions such as clicking, typing and scrolling, and a persistent shell with Python. Microsoft published it on 30 October 2026 under the Apache 2.0 license. The idea is that a small model should click only when it needs to see something, and write code or shell commands when a file edit is cheaper and easier to verify.
It starts from an intermediate 4B Mage-VL checkpoint (the Mage-ViT encoder plus Qwen3-4B-Instruct-2507) and goes through five training stages: midtraining on screen grounding, fine-tuning on GUI agent trajectories, fine-tuning on coding and terminal tasks, hybrid fine-tuning that teaches the choice between the two tools, and online reinforcement learning with checkable task rewards. Microsoft generated 293,806 training tasks across 6,915 skills with a workflow-first generator it calls Mage-CUA-Forge, and kept 318,837 verified trajectories, produced by a GPT-5.4 teacher whose tool harness was first pruned of overlapping tools. The model keeps the three latest screenshots and four full responses in view within a 32,768-token working window, although the architecture supports 262,144 tokens, and each turn contains a rationale, an action summary and the action itself.
Microsoft reports 66.0% on OSWorld-Verified with a 50-step budget, the average of two runs (65.33% and 66.75%), which it calls the best result among 4B models. In the same table a 9B model, ScaleCUA-Qwen3.5-9B, scores 68.7%, Muse Glimmer-30B is listed at 65.9%, Qwen3.6-27B at 63.9% and GPT-5.4-mini at 69.6%, and the closed frontier models are listed between 73.1% and 78.7%. Online reinforcement learning added about 10 points over the hybrid fine-tuned checkpoint while cutting model calls by 13% and output tokens by 25%. In a timed example against a 27B model that only uses the GUI, the hybrid 4B model finished a spreadsheet task in 39 seconds against 9 minutes 56 seconds. These are the developer's results. The model uses custom code, so load it with remote code enabled, and the repository includes the OSWorld agents and an SGLang Docker script.
Good for
- Prototyping a self-hosted computer-use agent for office tasks, such as documents, spreadsheets and presentations, on one GPU
- Linux desktop workflows that mix screen steps with file and terminal operations
- Research on training small computer-use agents, since the card describes the curriculum, data synthesis and harness design
- Reproducing OSWorld results, because the evaluation agents ship in the repository
Watch out for
- It runs shell commands and clicks for real, so run it in a sandbox or virtual machine
- The 66.0% score is below Microsoft's human reference of 72.4%, and a 9B model scores higher in the same table
- It uses a custom architecture and processor, a 32K working window and only the three latest screenshots
- The benchmark numbers are Microsoft's own, and the training trajectories came from proprietary teacher models such as GPT-5.4
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Mage-VL-Cowork (4B) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Mage-VL-Cowork (4B) BF16 | 4.74B | 11 GB | 5.45 GB | 3.27 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is Mage-VL-Cowork?
- Mage-VL-Cowork looks at a screenshot and operates a computer: it clicks, types, and runs shell commands and Python. Because it is only 4 billion parameters, it can run on a single GPU. It is free for commercial use under Apache 2.0, but it takes real actions, so run it in a sandbox.
- Can I use Mage-VL-Cowork commercially?
- Yes. Mage-VL-Cowork is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Mage-VL-Cowork open source?
- Mage-VL-Cowork is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Mage-VL-Cowork on my own computer?
- Yes. The smallest variant, Mage-VL-Cowork (4B), needs about 3.27 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Mage-VL-Cowork released?
- Mage-VL-Cowork was released by Microsoft on October 30, 2026.
- How long a context window does Mage-VL-Cowork have?
- Mage-VL-Cowork supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Mage-VL-Cowork model card (Hugging Face) huggingface.co
- Mage-VL-Cowork project page (Microsoft) microsoft.github.io
- Mage-VL-Cowork technical report (PDF, GitHub) github.com
- Mage repository (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Mage-VL Microsoft | 4.74B | 256K | Apache 2.0 (Permissive) | 3.27 GB Laptop or phone | |
| Qwen3-VL Alibaba Qwen | 2.13B – 236B 22B active | 256K | Apache 2.0 (Permissive) | 1.47 GB Laptop or phone | |
| Muse Glimmer Meta | 29.6B | 128K | Apache 2.0 (Permissive) | 20 GB Gaming GPU | |
| Needle 3 Cactus Compute | 121M | 8K | Apache 2.0 (Permissive) | 0.0835 GB Laptop or phone |