Microsoft Mage-VL-Cowork: a 4.7B-parameter hybrid agent for GUI and terminal

On October 30, 2026, the model microsoft/Mage-VL-Cowork was published on HuggingFace with the image-text-to-text task. It is a compact vision-language model trained to combine two channels of computer interaction: clicking through interfaces based on screenshots and executing code via the terminal. The authors emphasize that the model decides on its own when to act through the GUI and when to write code, without an external orchestrator.

What it means

The model has 4.7 billion parameters, the weights are published in safetensors format and take up 8.8 GB. The license is apache-2.0, and it runs out of the box via transformers. The hub already has two quantized builds — GGUF and MLX; the most compact working version weighs 2.5 GB in Q4_K_M quantization. That is a level that fits into the VRAM of most modern consumer GPUs. The author describes the training as a five-stage process starting from an intermediate Mage-VL checkpoint, with a synthetic dataset of 293,806 tasks. For local runs, the quantized builds are of practical interest: the full 8.8 GB weights will require a noticeable VRAM headroom, whereas the 2.5-gigabyte GGUF version lowers the entry threshold. More news on local AI tools is in the CPU3D news feed.