Mage-VL-Cowork Vision-Language Model: 12GB+ VRAM, GGUF Builds

At a Glance
| Item | Value |
|---|---|
| Repository | microsoft/Mage-VL-Cowork |
| Family guide | Mage-VL guide (1 articles) |
| Publisher guide | Microsoft: models and licenses |
| Published | 2026-10-30 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Microsoft has released “Mage-VL-Cowork", a hybrid agent that handles computer operations and code execution within a single vision-language model. Although it is a 4B (4-billion) parameter model, it can dynamically switch and execute GUI operations based on screen screenshots (computer_use) and terminal code/shell execution (bash) without requiring an external orchestrator.
It was developed based on the 4B checkpoint of the base model “Mage-VL“, through 5-stage curriculum learning and online reinforcement learning (RL). It aims to efficiently execute complex PC operation tasks in local and on-device environments by properly switching between screen layout operations and structured file editing and verification.
Specifications
- Parameters: 4B (4.7B)
- Architecture: llava_onevision2
- Context Length: 32,768
Performance
According to evaluation results on the OSWorld-Verified benchmark (50-step limit, 361 tasks) published by the creators, Mage-VL-Cowork records excellent scores for a 4B class model. Below is an excerpt of major models from the comparison table posted in the model card.
→ Scroll horizontally to see all columns
| Model | Params | Steps | OS | Office | Daily | Professional | Workflow | Overall |
|---|---|---|---|---|---|---|---|---|
| GPT-5.4-mini† | — | 50 | 79.2 | 68.0 | 76.6 | 69.4 | 63.3 | 69.6 |
| Claude Sonnet 4.5 | — | 100 | 70.8 | 72.6 | 61.4 | 63.2 | 49.5 | 62.9 |
| ScaleCUA-Qwen3.5-9B | 9B | 50 | 79.2 | 70.6 | 74.3 | 89.5 | 48.1 | 68.7 |
| Mage-VL-Cowork-4B (Ours) | 4B | 50 | 87.5 | 64.9 | 65.0 | 66.3 | 62.7 | 66.0 |
| Muse Glimmer-30B | 30B | — | — | — | — | — | — | 65.9 |
| Qwen3.6-27B | 27B | — | — | — | — | — | — | 63.9 |
| OpenCUA-72B | 72B | 50 | 45.8 | 42.7 | 53.8 | 75.5 | 22.3 | 44.9 |
| Fara1.5-4B | 4B | 50 | 25.0 | 7.6 | 29.5 | 12.2 | 6.6 | 13.9 |
*Some models have been omitted for better visibility. OSWorld is a metric (unit: success rate %) measuring whether tasks can be completed by operating an actual OS screen.
According to the model card report, this model records an overall score of 66.0%, with particularly high scores in OS operation tasks (OS) at 87.5% and multi-application processing (Workflow) at 62.7%. On the other hand, in items such as Office operations (64.9%), daily tasks (Daily: 65.0%), and professional tasks (Professional: 66.3%), it falls short of the larger parameter ScaleCUA-Qwen3.5-9B and GPT-5.4-mini. Still, it demonstrates overall superior results to models in the same scale range, such as Fara1.5-4B (13.9%), and even larger models like Qwen3.6-27B (63.9%).
Additionally, regarding the foundational performance of the base model Mage-VL (4B), validation data described in the model card shows scores of 95.14% on DocVQA for reading and answering document images such as invoices, 84.88% on ChartQA for reading numerical values from graph images, and 81.80 on OCRBench for recognizing text in images, equipping it with a high level of underlying vision and document comprehension capabilities.
Strengths and Use Cases
This model excels at executing tasks while dynamically routing between GUI operations (computer_use), which visually understand PC screen screenshots, and terminal shell command execution (bash) within a single model.
Specifically, it selects GUI operations in scenarios requiring visual layout confirmation on the screen, canvas editing, or live application manipulation. On the other hand, for tasks where shell utilization is more reliable and faster, such as structured editing of spreadsheets, documents, and presentation files, or file searches and deterministic verification of processing results, it automatically switches to command execution via bash.
This makes it possible to handle a diverse range of PC operation tasks on the OS in one stop, from web browser navigation and complex data processing spanning multiple applications to system settings changes, programming, and terminal work.
Furthermore, as foundational capabilities inherited from the base model Mage-VL, it features advanced visual comprehension abilities such as document image understanding for invoices and forms (DocVQA), reading graphs and charts (ChartQA), and text recognition within images (OCRBench), allowing it to perform operations with high precision regarding UI elements and text information on the screen.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 4.7B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 12GB (RTX 4070 / 3060 12GB, etc.) | BF16 | 8.8GB | 10.6GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: not registered, vLLM: not registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but ggml-org/Mage-VL-Cowork-GGUF provides a GGUF build you can use.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| tencent/Youtu-Parsing-Omni | 5.3B | 12GB | youtu-parsing | Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM (2026-11-17) |
| nvidia/Muse-Spark-Open-816B-A42B-DFlash2 | 4.4B | 12GB | nvidia-open-model-agreement-license | Muse-Spark-Open-816B-A42B-DFlash2 Vision-Language Model: 12GB+ VRAM (2026-11-17) |
| tencent/WeVisOne-VL-4B | 4.8B | 4GB | apache-2.0 | Tencent Releases WeVisOne-VL Vision Language Models (2026-11-17) |
| inclusionAI/ZwZ2-9B-GGUF | 9.4B | 8GB | apache-2.0 | ZwZ2-9B-GGUF Vision-Language Model: 8GB+ VRAM (2026-11-17) |
| inclusionAI/ZwZ2-4B-GGUF | 4.5B | 4GB | apache-2.0 | ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM (2026-11-17) |
How to Get It
This model is available in safetensors format on the Hugging Face microsoft/Mage-VL-Cowork repository. The license is Apache-2.0.
Since a custom architecture called llava_onevision2 is used for loading the model, trust_remote_code=True must be specified when loading via transformers or inference engines.
Libraries required for use can be installed with the following command:
pip install "transformers>=4.57.3" accelerate pillow torch torchvision qwen-vl-utils
Inference engines such as SGLang are supported, and it is also possible to start a server using the official Docker container image (xiaoyizhang/mage-vl-cowork:0.1.0) with the necessary dependencies already set up.
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:0.1.0
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-07 | ggml-org | GGUF | ggml-org/Mage-VL-Cowork-GGUF | Q4_K_M 3.0GB (fits in 4GB VRAM) |
File sizes of each build:
- Available builds in ggml-org/Mage-VL-Cowork-GGUF: Q4_K_M 2.5GB / Q8_0 4.4GB / BF16 8.2GB
In addition, 8 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- Nanbeige4.3-33B 33B Model for Coding and Agents: 12GB+ VRAM
- UserRepGen-4B Text Generation Model: 4GB+ VRAM, GGUF Builds
- ZoomGain-4B: 4GB+ VRAM, GGUF Builds
What to Read Next
- Find models by VRAM (This model runs from the 12GB tier) → Other models that run on a 12GB GPU
- Explore the same model family → Mage-VL family overview (1 articles, 1 converted builds)
- Engines that run this model → llama.cpp / Ollama
- Formats this model is available in → GGUF format guide and models
- Learn about the publisher → Microsoft: models, licenses and articles
Sources
Update History
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

