What is Mage-VL-Cowork?
Mage-VL-Cowork is a 4-billion-parameter computer-use agent from Microsoft's Mage team, published on Hugging Face on 30 October 2026 under Apache 2.0 with a technical report. It looks at screenshots and acts on a Linux desktop through two channels: a computer_use tool with 11 mouse and keyboard actions, and a bash tool with a persistent shell and Python environment. The model decides on each turn which to use, without a separate router.
It is built on Mage-VL, Microsoft's 4B vision-language model, which pairs the in-house Mage-ViT encoder with Alibaba's Qwen3-4B-Instruct-2507. Training ran in five stages, from screen grounding through GUI and coding fine-tuning to online reinforcement learning on 1,988 verifiable tasks, using 32 Nvidia B200 GPUs. Much of the training data came from a GPT-5.4 teacher whose tool set was pruned first; removing the accessibility-tree channel lifted the 4B student from 25.1% to 37.6% on Microsoft's held-out tasks.
Each turn returns a reasoning block, a one-line action summary and tool calls. The agent keeps the 3 most recent screenshots and folds older steps into one-line summaries to stay inside a 32,768-token window. Microsoft's GitHub repository says the Mage models are released for research purposes only.
Is Mage-VL-Cowork worth it?
Mage-VL-Cowork shows that a 4B model can do serious desktop work when it is allowed to write code as well as click. On OSWorld-Verified, averaged over two runs of 361 tasks with a 50-step budget, it scores 66.0%, against 65.9% reported for Muse Glimmer-30B, 63.9% for Qwen3.6-27B and 13.9% for the browser-focused Fara1.5-4B. The human reference is 72.4%. It is strongest on operating-system tasks at 87.5% and weakest on multi-application workflows at 62.7%.
The comparison needs care. Microsoft ran its own model; most rival figures were copied from their reports with different step budgets, and the model is not on the official OSWorld leaderboard. ScaleCUA-Qwen3.5-9B, at about twice the size, scores higher at 68.7%, and frontier models such as GPT-5.5 at 78.7% and Claude Opus 4.7 at 78.0% remain well ahead.
The hybrid channel is what makes it efficient. In Microsoft's LibreOffice Calc demonstration it merged cells, added a chart and saved the file in 6 steps and 39 seconds by editing the workbook in Python, where GUI-only Qwen3.6-27B took 39 steps and 9 minutes 56 seconds. After reinforcement learning it averaged 18.77 model calls and 5,499 output tokens per task, against 12,397 tokens for GPT-5.4-mini. On web tasks it is less dominant: an earlier checkpoint scored 58.6% on WebVoyager-Static against 68.0% for the browser-only Fara1.5-4B.
The conditions on its use matter more than the score. Apache 2.0 permits commercial use, but Microsoft's repository says the models are not intended for product or service deployment, and the card insists on a disposable sandbox because the agent runs arbitrary shell commands. The direct SGLang example also listens on all network interfaces. Testing covered Ubuntu desktops in English only, and the one quantised build is community work that stock llama.cpp cannot yet load.
Where it earns its place
An applied researcher at a university lab building desktop agents
Reproduce and extend a small computer-use agent on OSWorld
Use the bundled MAGEHybridAgent and the OSWorld harness in Microsoft's GitHub repository with the published settings: 50 steps, 3 screenshots kept, 1920 by 1080 resolution, temperature 0.7. The full training recipe is documented, but the task factory, training code and data are not public.
An automation engineer at a 300-person finance firm with Linux back-office machines
Prototype an agent that updates spreadsheets and reports
Run it in snapshot-reverted VMs with no access to credentials or the internal network, as the card advises, and let it edit .xlsx and .pptx files through Python in the shell. Microsoft's research-only statement means a production rollout needs a legal review first.
A developer with a 16 GB laptop GPU testing local agents
Run a desktop agent without a cloud API
The official BF16 weights are 9.48 GB, and a community Q4_K_M GGUF is 2.716 GB plus a 0.353 GB vision projector. That build needs a forked llama.cpp, because stock llama.cpp, Ollama and LM Studio cannot load its projector.
Before you commit
- Setup
- Pull the Docker image Microsoft documents (SGLang 0.5.2, PyTorch 2.8.0, Transformers 5.7.0) and mount the weights, or launch SGLang with trust_remote_code, a 32,768-token context and SGLANG_MAX_PIXELS=4000000. Use the verbatim hybrid system prompt from the card.
- Integrations
- SGLang's native /generate endpoint with base64 screenshots, Transformers with the bundled llava_onevision2 model code, and the MAGEHybridAgent harness for OSWorld. Not drop-in compatible with standard Qwen3-VL classes.
- Data and privacy
- Self-hosted only. Apache 2.0, but Microsoft's Responsible AI note says research only, with training data from public, licensed and internal sources. The card requires isolated sandboxes, no mounted credentials and network allow-listing against prompt injection.
- Cost at scale
- BF16 weights are 9.48 GB and one ModelFit calculation puts a full 32K-token cache at about 4.8 GB, so one 24 GB GPU holds both. Microsoft publishes no minimum hardware; each task averaged 18.77 model calls and 5,499 output tokens in its tests.
- Limits
- 32,768-token working window with 3 screenshots, about 2,000 tokens each at 1920 by 1080 by ModelFit's estimate. Up to 15 tool calls per turn; Linux and English only; the direct SGLang example binds to 0.0.0.0, while the Docker example binds to 127.0.0.1.
What’s good
- 66.0% on OSWorld-Verified at 4B, above Qwen3.6-27B (self-reported)
- Chooses between GUI actions and a bash shell on each turn
- 18.77 model calls and 5,499 output tokens per task after RL
- Apache 2.0 weights with a detailed training recipe and OSWorld harness
What’s not
- Microsoft says it is for research, not product deployment
- Tested only on English-language Ubuntu desktops
- Trails ScaleCUA-Qwen3.5-9B and frontier models on OSWorld
- Runs arbitrary shell commands; needs a sandbox and careful network setup
Best for
Researchers and agent-platform teams studying small computer-use agents on Linux, who can run them in disposable sandboxes.
Mage-VL-Cowork alternatives
Mage-VL-Cowork: common questions
- Is Mage-VL-Cowork worth it?
- Mage-VL-Cowork gets a 4B model within a few points of the best open desktop agents by letting it reach for a terminal, but Microsoft labels it research-only and Linux is all it has been tested on. We rate it 3.9 out of 5.
- How much does Mage-VL-Cowork cost?
- Free under Apache 2.0, though Microsoft's repository says the Mage models are for research and not intended for product or service deployment. No hosted API.
- What are the downsides of Mage-VL-Cowork?
- Microsoft says it is for research, not product deployment; tested only on English-language Ubuntu desktops; trails ScaleCUA-Qwen3.5-9B and frontier models on OSWorld; runs arbitrary shell commands; needs a sandbox and careful network setup.
- Who is Mage-VL-Cowork best for?
- Researchers and agent-platform teams studying small computer-use agents on Linux, who can run them in disposable sandboxes.
- How much does Mage-VL-Cowork cost for a team or in production?
- BF16 weights are 9.48 GB and one ModelFit calculation puts a full 32K-token cache at about 4.8 GB, so one 24 GB GPU holds both. Microsoft publishes no minimum hardware; each task averaged 18.77 model calls and 5,499 output tokens in its tests.
- Is Mage-VL-Cowork safe for company data?
- Self-hosted only. Apache 2.0, but Microsoft's Responsible AI note says research only, with training data from public, licensed and internal sources. The card requires isolated sandboxes, no mounted credentials and network allow-listing against prompt injection.
- What are the best Mage-VL-Cowork alternatives?
- The closest alternatives we have reviewed are Muse Glimmer, Qwen3.8-27B, MiniCPM-V 4.7 (35B-A3B and 1B) and Nemotron 3.5 Super VL.
Answers checked Nov 2, 2026.