Microsoft's Mage-VL-Cowork on a Mac: a 4B computer-use agent in MLX
NetProjects
Most computer-use models work from pixels alone: they look at a screenshot, move the mouse and type. Microsoft's Mage-VL-Cowork does that too, but it can also open a shell. It has two tools in one policy, computer_use for the mouse and keyboard and bash for a persistent terminal, and it learns when to click and when to write code. Microsoft reports 66.0% on OSWorld-Verified, the best result it knows of at the 4B scale.
It is a 4B model, a Mage-ViT vision encoder with a Qwen3-4B language model, released with open weights under the Apache-2.0 licence. We ported it to MLX so it runs on Macs with Apple Silicon. At 4B it is small enough to run on a laptop next to the desktop it operates. We checked every build against Microsoft's own code and published them openly on Hugging Face.
Always up to date. We keep these builds current on Hugging Face. When Microsoft releases a new version of Mage-VL-Cowork, we rebuild, re-test and publish it there, so the links on this page always point to the latest version. The numbers in this article were measured on the version named in each section; the model cards on Hugging Face carry the current figures.
- Hugging Face: https://huggingface.co/netprojects
- NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds
What the model does
You give it a screenshot, an instruction and the previous actions. It answers with a thinking block, a short Action line and one or more tool calls: a mouse or keyboard action on the screen, or a shell command. Coordinates are normalised to a 0 to 999 range on both axes, whatever the screen size, and the caller maps them to pixels. Microsoft's original is at https://huggingface.co/microsoft/Mage-VL-Cowork.
Microsoft's card reports 66.0% on OSWorld-Verified under a 50-step budget, averaged over two full runs, with 87.5% on operating-system tasks and 62.7% on multi-application workflows. These are Microsoft's figures, from its own harness; we have not reproduced them.
What we released
Four MLX builds on Hugging Face, converted from Microsoft's weights at revision 6fbe7f9. The quantised builds quantise the language model to affine with group 64 and keep the vision tower in bf16.
| Hugging Face | Size |
|---|---|
| netprojects/Mage-VL-Cowork-MLX-6bit (recommended) | 4.2 GB |
| netprojects/Mage-VL-Cowork-MLX-8bit | 5.3 GB |
| netprojects/Mage-VL-Cowork-MLX-4bit | 3.1 GB |
| netprojects/Mage-VL-Cowork-MLX-bf16 | 9.5 GB |
We tested the builds with mlx 0.32.4 and mlx-vlm 0.7.11. mlx-vlm 0.7.11 and newer have the mage_vl architecture built in, so these builds need no remote code.
How well it matches the original
The conversion only maps tensor names and config fields to mlx-vlm's layout; every tensor is copied unchanged. On a small random-weight model built from Microsoft's code, MLX and PyTorch logits agree to within 1e-6 on CPU. All weight files of the upstream revision match the Hub's SHA-256 hashes, and the chat template is the original, byte for byte.
We then compared each build with Microsoft's own PyTorch code, Transformers 4.57.3 in float32 on CPU, at the same pinned revision, on 16 agent first steps with screenshots (synthetic desktop screens and ScreenSpot-v2 screenshots) and Microsoft's system prompts.
Measured on Mage-VL-Cowork, 17 November 2026, MacBook Pro M5 Max (128 GB).
| Build | Same next token as the original | Mean NLL (original 0.413) |
|---|---|---|
| MLX bf16 | 98.6% | 0.414 |
| MLX 8-bit | 98.2% | 0.414 |
| MLX 6-bit | 96.5% | 0.418 |
| MLX 4-bit | 89.3% | 0.469 |
The token-parity figures are teacher-forced over 3,489 tokens on identical inputs. The torch-free mlx-vlm processor produces the same input tokens as Microsoft's processor on all 16 prompts. In every build the original chat template round-trips computer_use and bash calls and tool responses. On 8 first-step prompts with Microsoft's hybrid system prompt, 7 answers per build are well-formed tool-call JSON with valid actions and coordinates in range; the eighth was still thinking at our 320-token test limit.
The 6-bit build is the one we recommend for its balance of size and fidelity. The 8-bit and bf16 builds track Microsoft's model most closely. The 4-bit build is the smallest and agrees on 89.3% of tokens, so expect more drift in its reasoning and coordinates.
Try it
On a Mac with Apple Silicon, with mlx-vlm:
pip install -U "mlx-vlm>=0.7.11"
mlx_vlm.generate --model netprojects/Mage-VL-Cowork-MLX-6bit --max-tokens 1024 \
--image screenshot.png --prompt "Turn on Dark Mode."For agent use, send Microsoft's exact system prompt and the first-step user turn format from its evaluation harness, which are described on the model card. To serve it as a local endpoint for an agent harness:
mlx_vlm.server --model netprojects/Mage-VL-Cowork-MLX-6bit --port 8100Screenshots up to 4,000,000 pixels are used at full resolution. To map a coordinate to pixels, use x_px = round(x / 999 * (width - 1)), and the same for y with the height, as Microsoft's agent does. The model card has the full commands: https://huggingface.co/netprojects/Mage-VL-Cowork-MLX-6bit
Use it in a sandbox
Mage-VL-Cowork is an autonomous agent that can run arbitrary shell commands and GUI actions. Follow Microsoft's guidance. Run it in an isolated container or VM with no access to host files, local networks or personal environment variables. Require explicit human confirmation before irreversible actions such as external submissions, signing in, payments or privilege changes. Do not mount SSH keys, cloud tokens, browser cookies or password stores. Restrict outbound network access when it reads untrusted web content, to limit prompt injection hidden in pages or documents.
Honest limits
- All figures come from one machine.
- We checked parity on 16 first steps and 8 tool-call prompts. We did not run OSWorld-Verified or any other agent benchmark, so we cannot say how the builds score at the full task.
- We have no speed figures yet, and no memory figures beyond the file sizes above.
- We have not run the model on a live desktop.
- The 4-bit build differs more from the original than the others.
Licence
Mage-VL-Cowork, Qwen3-4B and our builds are all under the Apache-2.0 licence. Each of our repositories carries the licence text and a notice that describes the conversion and credits Microsoft and Alibaba Cloud (Qwen). If you use the model, Microsoft asks you to cite its paper; the citation is on the original model card.
Links
- MLX 6-bit build (recommended): https://huggingface.co/netprojects/Mage-VL-Cowork-MLX-6bit
- MLX 8-bit build: https://huggingface.co/netprojects/Mage-VL-Cowork-MLX-8bit
- MLX 4-bit build: https://huggingface.co/netprojects/Mage-VL-Cowork-MLX-4bit
- MLX bf16 build: https://huggingface.co/netprojects/Mage-VL-Cowork-MLX-bf16
- NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds
- Microsoft's original: https://huggingface.co/microsoft/Mage-VL-Cowork
- All our AI ports: https://netprojects.com/ai
Always up to date. We keep these builds current on Hugging Face. When Microsoft releases a new version of Mage-VL-Cowork, we rebuild, re-test and publish it there, so the links on this page always point to the latest version. The numbers in this article were measured on the version named in each section; the model cards on Hugging Face carry the current figures.
- Hugging Face: https://huggingface.co/netprojects
- NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds
Credits
Mage-VL-Cowork is by Microsoft, built on Qwen3-4B by Alibaba Cloud's Qwen team, under the Apache-2.0 licence. The MLX builds are by NetProjects (Internet Projects Ltd). Not affiliated with Microsoft or Alibaba Cloud.
This article was drafted with AI assistance, then edited and approved by people on our team. You can read more about how we use AI at https://netprojects.com/standards/ai.