A Guided Tour of the New Microsoft Foundry Labs

Every developer has that exhilarating moment when they manage to bring a bold idea to life on their screen, solving their specific problem. Microsoft Foundry Labs is here to help you relive that moment over and over again.

Imagine a place where Microsoft’s cutting-edge AI research transforms into tools you can engage with right away. Whether you’re interested in computer usage, predicting biomolecular structures, generating 3D visuals in real-time, or exploring multilingual speech, that technology is accessible. No more waiting for papers to be published or for permission to access these innovations. They’re interactive and just a click away for your next project. Let’s dive into what’s available for you.

At Foundry Labs, there’s a simple belief at its core, expressed on their About page: AI is progressing rapidly, and so should our access to it. Traditionally, there’s been a frustrating lag time between groundbreaking research and practical application — often taking months or even years for concepts to become tools. Foundry Labs is designed to bridge that gap, delivering experiments straight from Microsoft’s labs directly to developers and researchers like you, centred around three main actions: discover cutting-edge innovations, experiment with them, and connect with others who are advancing the field.

The research is not scattered randomly; it’s focused on six key areas where AI is making a significant difference.

Select a domain that resonates with your work, and Labs will guide you through the forefront of innovation in your field.

The homepage greets you with the phrase “Breakthrough AI you can try today.” It showcases a rotating selection of experiments worth your attention, each featuring a Try it now button to jump directly into Microsoft Foundry. It also includes a Learn more option for those wanting to delve deeper. Here’s a sneak peek at some highlights:

• MAI-Image-2.5 — Microsoft AI’s latest image-generation model, featuring image-to-image editing capabilities and exceptional control, recently ranking No. 2 in editing performance on Arena.ai.
• MAI-Thinking-1 — This is Microsoft AI’s initial large language model, designed for enhanced reasoning and mathematical functions at a reduced cost.
• MagenticLite — An open-source app tailored for smaller models, succeeding Magentic-UI and combining the MagenticBrain orchestrator with Fara 1.5 for seamless operation on your hardware.
• Fara 1.5 — These small agentic language models can interpret screen information from screenshots and anticipate user clicks.
• EO/OS Object Detection — A first-party Earth Observation model adept at detecting and localising objects within satellite and aerial images on a petabyte scale, contributing to the GeoAI domain alongside the Planetary Computer team.

As you scroll down, keep an eye on the “Just added to Labs” section, which highlights the latest arrivals like VibeVoice ASR, MAI-Image-2-Efficient, Magentic Marketplace, and BugPilot. It’s the quickest route to discover recent additions.

The homepage is a preview; the Innovations tab is the treasure trove — containing over 50 experiments that continue to expand, drawn from the latest research and available for you to use today. The visuals throughout the site are crafted using Microsoft’s proprietary MAI image models.

This section is where it gets truly engaging. You can filter and sort the entire catalogue to quickly find what’s relevant to you — whether you’re searching by the six domains, the latest additions, or specific artifact types. Looking for a vision model in Creative & Generative Media? A ready-to-use framework in Code & Software Engineering? With just a few clicks, you’ll have your personalised shortlist. Each result opens with a dedicated page that explains the experiment in detail and allows you to try it out.

For example, take a look at Trellis — it’s an excellent place to begin, not only because it outlines the model but also because it enables you to use it immediately.

An interactive playground, right on the page. TRELLIS transforms a single image (or a text prompt) into a fully-textured 3D asset with a lively interface for experimentation. You can upload an image, adjust parameters—like latent and sparse-structure CFG scale, sampling steps, and seed—then click Generate 3D Model to preview your creation. Once satisfied, you can export a GLB file and integrate it directly into your workflow. There’s no setup required, no need for a separate notebook, and you don’t need your own GPU.

The details behind the demo. Below the interactive section, the page goes further by explaining what makes TRELLIS so innovative. It uses a unique Structured LATent (SLat) representation combined with rectified-flow transformers scaled to 2 billion parameters. Trained on 500,000 diverse 3D objects, it can deliver meshes, radiance fields, and 3D Gaussians — allowing you to choose the output format essential for your application without needing retraining. It takes less than 10 seconds to convert an image to a textured mesh on a single A100 GPU and supports local edits in specific 3D areas without affecting the rest. It was later adopted by NVIDIA AI Blueprints in September 2025. The usage stats on the page, exceeding 2.4 million users, indicate its popularity.

Everything you need for further exploration. Clear tags categorise it (Creative & Generative Media · Model · Vision · Experimental), the technology stack is outlined (PyTorch, diffusion models, NeRF, 3D Gaussians, CUDA), and a “Ready to Explore?” section links you to the GitHub repository, the research article, and the project blog. That’s the complete journey on one page: try it out, grasp the concept, then take the code for your use.

Each experiment page follows this structure, so after exploring TRELLIS, you’ll be well-equipped to navigate any of them.

Testing a model is exhilarating; turning it into a product requires solid proof. This is exactly what the Stories page offers — real-life examples from teams that have successfully integrated Labs experiments into their workflows.

• Space Intelligence — A UK organisation focused on mapping the world’s forests. By integrating Microsoft Foundry with Microsoft Planetary Computer, they increased data production by 100 times and reduced mapping times from six months to just six weeks, covering 3 billion hectares across over 50 countries in one year.
• Sight Machine — Achieved a 10% boost in manufacturing productivity through Foundry-powered solutions.
• Commerzbank — Managed to scale customer conversations to 30,000 per month using Foundry Agent Service.
• MediaTek — Realised a 50% increase in on-device AI performance leveraging Phi models.

If you’re trying to present a case for adoption, these provide tangible evidence of effectiveness. And if you’ve launched something through Labs, you’re encouraged to share your own story for potential feature.

Excellent tools foster communities, and the Community page is a hub for over 25,000 developers and researchers who unite under the motto: build together, think further. It encompasses three key elements:

• An events calendar — Featuring appearances by the Foundry team and community at major events like Microsoft Build, Ignite, KubeCon, GitHub Universe, and NeurIPS, complete with links for registration.
• The latest from the labs — A stream of blogs and publications authored by the teams behind the experiments, keeping you updated as research evolves.
• The conversation — Direct access to platforms like Discord, Reddit, the Microsoft Research Blog, and Tech Community, enabling you to ask questions, share ideas, and watch as your suggestions influence upcoming developments.

Foundry Labs serves as the quickest route from Microsoft’s leading research to practical application. The frontier is open, the experiments are live, and the community is already thriving. Dive in and explore — and share what you create.