Mage: a family of multimodal models
A family of multimodal models from Microsoft, built at a fixed 4B-parameter budget.
Verified
Core facts
Read by a script from the official API, not rewritten by a model.
- Repository
- https://github.com/microsoft/Mage
- License
- Apache-2.0 (Mage-VL, Mage-VL-Cowork) / MIT (Mage-ViT, Mage-Flow)
- Languages
- Python
- Maintainer
- microsoft
- Verified
What it is
Mage is a family of multimodal models from Microsoft, built at a fixed 4B-parameter budget: Mage-VL for image and video understanding, Mage-Flow for image generation and editing, and Mage-VL-Cowork for hybrid computer use with GUI, code, and terminal actions. Mage is designed to make visual understanding, generation, and computer use accessible for controlled experiments, post-training research, and vertical-domain applications under realistic compute budgets; the project also states that the models are released for research purposes only and are not intended for product or service deployment. The family is organized around a shared codec-aligned efficiency philosophy, applied to both the understanding and the generation side. Mage-VL is a codec-native, proactive-streaming multimodal foundation model for image and video understanding, whose visual encoder Mage-ViT is trained entirely from scratch and paired with a Qwen3-4B decoder at 4B scale.
Main features
A single released Mage-VL checkpoint provides image and video understanding and the proactive streaming gate, as one model with no separate variants.
Mage-VL-Cowork is a 4B hybrid computer-use agent that combines visual GUI interaction with code and terminal execution in one policy; it starts from an intermediate Mage-VL checkpoint and uses a five-stage curriculum across midtraining, SFT, and RL.
The project reports that Mage-VL-Cowork reaches 66.0% on OSWorld-Verified under a 50-step budget at 4B scale.
Mage-Flow is a 4B generative stack for text-to-image generation and instruction-based image editing, built from Mage-VAE and a Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching; each task ships in Base, RL-aligned, and 4-step Turbo variants.
Mage-Flow-Edit supports semantic content editing, appearance transformation, image restoration, and structure-aware outputs within a unified image-and-text-conditioned model.
The project reports, at 1024² on a single A100, Mage-Flow-Turbo at 0.59 s/image and Mage-Flow-Edit-Turbo at 1.02 s/edit.
Getting started
Each model is self-contained in its own directory with a dedicated README.
The Mage-VL README covers details, installation, inference, and proactive streaming.
The Mage-VL-Cowork README covers an overview, the training recipe, and results.
The Mage-Flow README covers details, installation, Python API, CLI, and Gradio app.
Each Mage-Flow checkpoint is a self-contained diffusers-style repo.
Who it is for
The project states that the models are released for research purposes only and are not intended for product or service deployment. The models should be used in controlled research settings with appropriate human oversight, and downstream users are responsible for applying additional safeguards such as content moderation, validation, and compliance checks before broader use. Under certain prompts, the models may generate outputs that are inaccurate, biased, or inappropriate.
The training data includes public, licensed, and internal datasets that were processed to remove clearly identifiable personal information and reduce harmful content where possible. Mage-VL is licensed under Apache-2.0, and Mage-VL-Cowork is also licensed under Apache-2.0. Mage-ViT is licensed under MIT, and Mage-Flow is also licensed under MIT.
Related tools
Related tools listed on this site.
百灵大模型
Ant Ling is a language and multimodal model family. Official documentation describes online multi-turn conversations, text generation/content creation, API access and cross-modal understanding/generation.
MachineLearningMastery
A machine-learning learning-resource site for developers. Its public page offers topic guides, tutorials and ebooks covering Python, deep learning, data preparation and related mathematics.
FlagEval
Evaluation toolkit and open platform for large models. Its public project covers foundation models, training and fine-tuning algorithms, with language, vision, audio and multimodal evaluation directions.
OpenCompass
Model evaluation platform with language and multimodal leaderboards, benchmark resources, and documentation.
MMBench
An OpenCompass Community multimodal-understanding benchmark using multiple-choice questions and fine-grained ability dimensions to analyze vision-language models, with Chinese and English data and evaluation-tool links.
CleverSee
Alibaba Cloud’s enterprise AI search and question-answering platform, describing web retrieval, private multimodal data and configurable applications.
01Agent
AI creation workspace for marketing teams. Its site describes articles, visual posts, posters, presentations, reports, e-commerce assets and videos from briefs, with further editing on an infinite canvas and reusable skill templates.
魔珐星云
Xingyun is an open digital-human and embodied-interaction platform describing perception, knowledge, speech/action expression, device-side rendering and real-time SDKs.
Vivix
Real-time interactive multimodal platform describing responsive characters, scenarios and worlds with model, research and developer entry points.
LightX2V Studio
LightX2V Studio’s public page positions it as an AI digital-human video and multimodal-content platform, starting from character assets, narration scripts and AI voices and organizing work through projects, Canvas and reusable workflows.
Sources
This page is written from the official sources below. Open them to check the original text.
- GitHub READMEhttps://github.com/microsoft/MageFetched
Verified ·
This page is written by AI from the official sources above and checked automatically. This site has not tested it. Found a mistake? Tell us.