Last updated: November 6, 2026
Mage-VL-Cowork is Microsoft Mage Team’s 4B hybrid computer-use agent that chooses GUI clicks or terminal code inside one policy. Released around October 30, 2026 under Apache-2.0 on Hugging Face, it scores 66.0% on OSWorld-Verified with a 50-step budget—the strongest open result at that size.
This guide summarizes the official project page, Hugging Face model card, and arXiv:2610.40747 technical report. For other open computer-use stacks, see our NorthCUA guide and Gemini Computer Use guide.
Contents
- What Mage-VL-Cowork is
- Specs and action channels
- OSWorld-Verified and WebVoyager results
- How to install and serve safely
- Five-stage training recipe
- How it compares with larger CUAs
- Limitations and safety
- Frequently asked questions
What Mage-VL-Cowork is
Effective computer-use agents should mix visual GUI control with code and shell execution. Mage-VL-Cowork does that in a single 4B vision-language policy rather than an external GUI-versus-coding orchestrator. The Microsoft Mage Team trains it from an intermediate Mage-VL checkpoint through a progressive five-stage curriculum across midtraining, supervised fine-tuning, and online reinforcement learning.
Upstream of training, the team evolves the teacher harness (tool interfaces, prompts, and preinstalled libraries) and builds Mage-CUA-Forge, a workflow-first synthesizer that produced 293,806 desktop and web tasks across 6,915 skills while cutting recurring per-task generation cost by 82.6% ($0.906 to $0.158). Retained hybrid trajectories number 318,837.
If you are comparing self-hosted coworker agents rather than pure CUAs, our Open Duet 0.8 guide covers a different packaging model. For coding-first agents, see SWE-2 Flash and Claude Code 2.1.318.
Specs and action channels
The model card lists the unified checkpoint microsoft/Mage-VL-Cowork with a custom llava_onevision2 architecture (Mage-ViT encoder plus Qwen3-4B). Always load with trust_remote_code=True.
| Property | Specification |
|---|---|
| Developer | Microsoft Mage Team |
| Parameters | 4B (about 9.48 GB bfloat16 across 2 safetensors shards) |
| Context | 32,768 active tokens (architecture supports up to 262,144) |
| Working memory | 3 most recent screenshots + 4 full responses + folded Action history |
| GUI channel | computer_use with 11 primitives; coordinates normalized to [0, 999] |
| Terminal channel | bash persistent shell and Python environment |
| Turn format | rationale + Action summary + tool_call blocks |
| License | Apache-2.0 |
Recommended sampling for the reported OSWorld configuration is temperature 0.7 (or 0.0 for deterministic runs), top_p 0.8, and max_new_tokens 4096.
OSWorld-Verified and WebVoyager results
On 361 held-out OSWorld-Verified tasks excluding Google Drive, with a 50-step budget, Mage-VL-Cowork averages 66.0% across two runs (65.33% and 66.75%). Category scores include OS 87.5%, Office 64.9%, Daily 65.0%, Professional 66.3%, and Workflow 62.7%.
| Model | Size | Steps | Overall |
|---|---|---|---|
| ScaleCUA-Qwen3.5-9B | 9B | 50 | 68.7% |
| Mage-VL-Cowork | 4B | 50 | 66.0% |
| Muse Glimmer-30B | 30B | — | 65.9% |
| Qwen3.6-27B | 27B | — | 63.9% |
| EvoCUA-32B | 32B | — | 56.7% |
| OpenCUA-72B | 72B | 50 | 44.9% |
| Fara1.5-4B | 4B | 50 | 13.9% |
Before RL, the Hybrid SFT checkpoint reaches 58.6% on WebVoyager-Static (377 of 643 tasks), second behind browser-specialist Fara1.5-4B at 68.0% and ahead of Qwen3-VL-8B-Thinking (53.7%) and Qwen3.5-9B (53.3%).
Online hybrid RL adds 10.16 percentage points over Hybrid SFT while cutting mean model calls by 13.0% (21.57 to 18.77) and output tokens by 25.4% (7,374 to 5,499 per task).
How to install and serve safely
- Download weights. Pull
microsoft/Mage-VL-Coworkfrom Hugging Face into a local directory under Apache-2.0. - Prefer the pinned Docker runtime. Pull
xiaoyizhang/mage-vl-cowork:0.1.0(SGLang 0.5.2, PyTorch 2.8.0+cu128, Transformers 5.7.0, FlashInfer 0.3.1) as documented on the model card. - Bind-mount the checkpoint read-only and publish only
127.0.0.1:30001so the OpenAI-style generate endpoint stays local. - Set serving flags. Use
--trust-remote-code,--dtype bfloat16,--context-length 32768, andSGLANG_MAX_PIXELS=4000000so 1920×1080 screenshots keep fidelity. - Keep the agent in a disposable sandbox. Run guest desktops in Docker or snapshot VMs with no host mounts, no personal credentials, and outbound allow-lists for untrusted pages.
- Wire the hybrid system prompt. Use the model card’s two-tool contract (
computer_useandbash) and the three-part turn format for best behavior.
export MODEL_DIR=/path/to/Mage-VL-Cowork
docker pull xiaoyizhang/mage-vl-cowork:0.1.0
docker run --rm --name mage-vl-cowork \
--gpus 'device=0' --shm-size=8g \
-p 127.0.0.1:30001:30001 \
--mount "type=bind,src=${MODEL_DIR},dst=/models/Mage-VL-Cowork,readonly" \
xiaoyizhang/mage-vl-cowork:0.1.0For general SGLang serving upgrades, see our SGLang 0.5.23 guide. CLI-oriented automation stacks such as Antigravity CLI solve a different workflow layer than this hybrid CUA.
Five-stage training recipe
Training starts from an intermediate Mage-VL 4B checkpoint and follows a specialize-then-unify curriculum:
- GUI-oriented midtraining for grounding, reasoning, and tool syntax (OSWorld-G 30.6% → 56.9%; ScreenSpot-v2 78.8% → 87.2%).
- GUI agent SFT on multi-step screenshot workflows (45.56% OSWorld-Verified).
- Coding and terminal SFT on repository and desktop programmatic edits.
- Hybrid CUA SFT on 318.8K hybrid trajectories (55.84%).
- Online hybrid RL with GRPO, asynchronous desktop rollouts, and verifiable rewards (66.0%).
Harness ablations show that pruning redundant accessibility routes before distillation matters: on 672 held-out multi-application tasks, a GUI+coding teacher at 92.0% yields a 4B student at 37.6%, versus 25.1% when accessibility is added—a 12.5-point student gap.
How it compares with larger CUAs
Mage-VL-Cowork is aimed at on-device and compact deployment rather than matching frontier proprietary CUAs such as GPT-5.5 (78.7%) or Claude Opus 4.7 (78.0%) on OSWorld-Verified. Against open-weight peers it is unusually strong at 4B: it trails ScaleCUA-Qwen3.5-9B overall but leads that 9B model on OS (87.5% vs 79.2%) and Workflow (62.7% vs 48.1%). Side-by-side demos on the project page show the hybrid policy finishing LibreOffice Calc chart tasks in 6 steps versus 39 GUI-only steps for Qwen3.6-27B.
Limitations and safety
- Custom architecture requires
trust_remote_code; it is not a drop-in for stock Qwen3-VL loaders. - Designed for isolated sandboxes; bash can mutate files and run commands.
- OSWorld headline uses a 50-step budget on 361 tasks excluding Google Drive; other budgets or suites will differ.
- WebVoyager-Static 58.6% is the SFT checkpoint before RL; live web scores are not claimed for the final RL model.
- Serving needs a capable GPU; the reference RL fleet used B200-class hardware, which is not a consumer requirement statement but signals serious multimodal load.
Frequently asked questions
What is Mage-VL-Cowork?
Mage-VL-Cowork is a 4B hybrid computer-use agent from the Microsoft Mage Team. One vision-language policy combines screenshot-grounded GUI actions with code and terminal execution, and scores 66.0% on OSWorld-Verified under a 50-step budget.
How does the hybrid GUI and bash channel routing work?
Each turn the same policy chooses computer_use for mouse and keyboard actions on the screen, or bash for shell and Python edits that return text. There is no external orchestrator; channel choice is learned inside the model.
Where can I download Mage-VL-Cowork and what is the license?
Weights are on Hugging Face as microsoft/Mage-VL-Cowork under the Apache-2.0 license. Evaluation and serving helpers live in the mage_vl_cowork directory of the microsoft/Mage GitHub repository.
How do I serve Mage-VL-Cowork with SGLang?
Microsoft documents a pinned Docker image xiaoyizhang/mage-vl-cowork:0.1.0 that runs SGLang with native llava_onevision2 support. Mount the checkpoint, expose port 30001 on localhost, pass trust_remote_code, and use bfloat16 with a 32,768-token context.
Is Mage-VL-Cowork safe to run on a personal machine?
No. The model card says it can run arbitrary bash and GUI actions and should stay in an isolated Docker container or snapshot-reverted VM. Do not mount SSH keys, cloud tokens, cookies, or password stores, and keep irreversible actions behind human confirmation.
How does Mage-VL-Cowork compare with larger computer-use models?
At 4B parameters and a 50-step budget it scores 66.0% on OSWorld-Verified, ahead of Muse Glimmer-30B (65.9%), Qwen3.6-27B (63.9%), EvoCUA-32B (56.7%), and OpenCUA-72B (44.9%). ScaleCUA-Qwen3.5-9B is higher overall at 68.7%, while Mage-VL-Cowork leads on OS (87.5%) and Workflow (62.7%). Frontier proprietary models remain ahead.
Sources
- Microsoft Mage: Mage-VL-Cowork project page
- Hugging Face: microsoft/Mage-VL-Cowork model card
- arXiv:2610.40747 Mage-VL-Cowork technical report
- GitHub: microsoft/Mage mage_vl_cowork directory
Image: Official Mage-VL-Cowork teaser from the Microsoft Mage project page (teaser-20261030-wx.png), resized to 800px JPEG for editorial use. Microsoft and Mage are trademarks of their owners. FreeAnswers is not affiliated with or endorsed by Microsoft.