Mage-VL-Cowork (Microsoft)

What is Mage-VL-Cowork (Microsoft)

Mage-VL-Cowork is a 5B-parameter vision model from Microsoft for building computer-use agents that operate both graphical interfaces and terminals. The weights are openly available for free under the Apache-2.0 license on Hugging Face, where the model was created on 2026-10-30.

How to use

1. Install dependencies: pip install "transformers>=4.57.3" accelerate pillow torch torchvision qwen-vl-utils Download microsoft/Mage-VL-Cowork from Hugging Face;
2. always pass trust_remote_code=True because it uses a custom llava_onevision2 architecture.
3. Serve it on a GPU with the provided Docker image (xiaoyizhang/mage-vl-cowork:0.1.0) or launch SGLang directly with --trust-remote-code, bfloat16 and a 32,768-token context.
4. Use the exact MAGE_HYBRID_SYSTEM_PROMPT and two-channel tool schema (computer_use and bash) from the model card.
5. Send desktop screenshots and task instructions to the server (SGLang POST /generate);
6. recommended sampling: temperature 0.7, top_p 0.8, max_new_tokens 4096.
7. Optionally run offline inference with Transformers or reproduce the OSWorld-Verified evaluation with the bundled osworld_eval scripts.

Use cases

• Building computer-use agents that combine GUI clicks and terminal/code execution in one 4B model Automating desktop tasks across operating-system and multi-application workflows Programmatically editing spreadsheets, documents and presentations via bash while using the GUI for visual tasks Research on on-device computer-use agents and reproducing OSWorld-Verified results

Price from official sources

Free

• Open weights released under the Apache 2.0 license.

Check the official site for current pricing · Price not yet checked by our editors

Source: huggingface.co · Info as of Nov 2, 2026

Company

Name

Microsoft

Social

• github