Mage-VL-Cowork MLX 4-bit

A 4-bit MLX build of microsoft/Mage-VL-Cowork, Microsoft's 4B computer-use model (Mage-ViT vision encoder + Qwen3-4B language model), so you can run it on an Apple Silicon Mac.

This is a community MLX conversion, not an official Microsoft release. All credit for the model goes to Microsoft. License stays Apache-2.0 (see LICENSE, copied from the original repo).

How it was made

1. The original checkpoint uses a different weight key layout than mlx-vlm expects, so the weight keys were remapped to the mage_vl layout. Only key names and a few config fields changed; the tensors were copied unchanged.
2. Converted with mlx_vlm.convert (mlx 0.32.4, mlx-vlm 0.7.11):
   • language model: 4-bit affine quantization, group size 64
   • vision tower: kept in BF16
   • about 3.14 GB on disk, 5.30 bits per weight on average

The remap script is in the benchmark dataset (raw/scripts/remap.py) if you want to redo it yourself.

How it loads

config.json has model_type: mage_vl, which is a built-in model in mlx-vlm 0.7.11 (mlx_vlm/models/mage_vl). No custom modeling code or trust_remote_code is needed. It was tested with stock mlx-vlm 0.7.11 on a 64GB M5 Pro Mac mini; older mlx-vlm versions without mage_vl won't load it.

pip install -U mlx-vlm

python -m mlx_vlm.generate \
  --model pratikgx/Mage-VL-Cowork-MLX-4bit \
  --image screenshot.png \
  --prompt "Turn on Dark Mode." \
  --max-tokens 512

For agent use, pass Mage's official system prompt and first-step message format from the original model card (--system works for the system prompt). The exact prompt used in the benchmarks is in the dataset at raw/scripts/system_prompt.txt.

Speed on a 64GB M5 Pro Mac mini

Measured with mlx-vlm 0.7.11, greedy, 256 generated tokens, median of 3 runs after 1 warmup. A background inference app was running on the Mac during the benchmarks, so an idle machine may be a bit faster.

One full agent turn (screenshot in, tool call out): 2.36s at 1280x800 and 3.10s at 1920x1080.

Caveats

• Speed only. No quality suite was run, and no OSWorld run.
• In a tiny 3-task demo, this 4-bit build missed a Dark Mode toggle click by 42px, while the BF16 MLX build got it right. That's a very small sample and not an OSWorld score, so don't read too much into it.
• Community port: if something looks off compared to Microsoft's PyTorch code, it's most likely the conversion, not the model.

Full results, raw logs, scripts and the demo traces: https://huggingface.co/datasets/pratikgx/mage-vl-cowork-m5-pro-mac-mini-bench

Write-up on GitHub

How I ran this, what broke, and the scripts: Microsoft's Mage-VL-Cowork 4B on my Mac mini. It's part of Local AI field notes, where I log every model I run on my Mac mini.