Microsoft researchers have recently unveiled an innovative foundation model capable of performing agentic functions.
Named Magma, this multimodal AI model can understand both images and text in various digital and physical contexts. This groundbreaking model has been pre-trained on extensive datasets that include text, images, videos, and spatial formats. According to the tech giant, Magma builds upon vision-language (VL) models, enabling it to not only interpret multimodal information but also to plan and take action based on that understanding. The AI agent-enabled model is versatile and suitable for tasks such as computer vision, user interface (UI) navigation, and robot manipulation.
While traditional VL models primarily focus on pairing images with text, they often fall short in understanding spatial relationships and executing actions. Magma enhances this capability by incorporating spatial intelligence, which allows it to predict movements, track objects, and carry out commands based on both textual and visual data.
Developed collaboratively by researchers from Microsoft Research, the University of Maryland, the University of Wisconsin-Madison, KAIST, and the University of Washington, Magma stands out as the first foundation model that can interpret and ground multimodal inputs in its environment.
In a post on GitHub, Microsoft researchers provided insights into the new Magma foundation model. Foundation models are unique large language models (LLMs) created from the ground up, rather than being derived from existing models. They often serve as the foundational basis for subsequent models in the series. What sets Magma apart is its pre-training on a diverse array of datasets.
The researchers noted that the underlying architecture of Magma is based on the Llama 3 AI model. However, Magma also possesses the capability to plan and act within the visual-spatial realm, enabling it to generate outputs like a chatbot while also executing physical actions.
Microsoft researchers have shared the benchmark scores of their AI model based on internal testing. It has shown impressive performance in all agentic evaluation tests, surpassing models from OpenAI, Alibaba, and Google. Currently, the company has not made Magma available to the public.
- Meta’s AI Video Editor Reimagines Short Videos Instantly - June 13, 2025
- Google Search Gets Smarter with Real-Time Data Visualisation Tools - June 9, 2025
- Bing’s New AI Video Creator Uses OpenAI’s Sora – And It’s Free - June 3, 2025
