At Microsoft Research, we are using AI to help developers expand their capabilities, streamline their work, and transform ideas into prototypes.
The projects Microsoft Research is featuring at BUILD are just a few of the many resources that MSR makes available to customers, partners, and other developers. You can check out many more MSR open-source technologies on Microsoft Foundry (opens in new tab) and GitHub (opens in new tab).
Here are a few examples of Microsoft Research technologies that show how AI is accelerating innovation and helping developers create next-generation products and services.
Hands-on models
Aurora
Traditional weather forecasts depend on supercomputers running for hours. Aurora, a foundation model that was trained on more than a million hours of atmospheric data, delivers state‑of‑the‑art forecasting with substantial gains in both speed and accuracy. Aurora generates predictions in seconds—around 5,000 times faster than traditional numerical models—while outperforming existing approaches on 91% of evaluated targets. Aurora goes far beyond traditional weather forecasts of short-range temperature and precipitation changes by enabling the prediction of air pollution, extreme storms, medium-range weather, ocean wave action, atmospheric chemistry, and regional climate shifts that were previously too expensive to forecast at all.
TRELLIS
It takes a skilled artist hours to create a production-quality 3D asset (textured, lit, and topology-correct). TRELLIS, a 4B-parameter 3D generative model that turns text or images into production-ready assets, can generate one from a text description or a single photograph (2D image) in seconds, with full physically based rendering (PBR) materials, arbitrary topology, and resolution up to 1536³ voxels. The TRELLIS-generated 3D asset can then be previewed, refined, and exported for downstream creative or technical workflows.
Promptions
Every AI image generator has the same bottleneck: the prompt box. You know roughly what you want to create, but getting from a vague idea to a specific visual means trial-and-error rewording and dozens of attempts to nudge style, composition, or mood. At best, the process is tedious and opaque. Promptions (“prompt” + “options”) eliminates that friction by inserting a middleware layer between the user and the model that surfaces the implicit choices buried in every prompt and replaces trial-and-error rewording with contextual UI controls. Adjust a toggle, and the image regenerates, so you steer the output by clicking, not rewriting.
Demo station experiences
MagenticLite
MagenticLite is the next generation of Magentic-UI, an agentic experience that works across the browser and local file system, now optimized for small models. It combines a redesigned application with a harness rebuilt for small language models, and ships alongside two purpose-built models—MagenticBrain for orchestration and Fara1.5 for computer use—codesigned to work as a single system.
OptiGuide/OptiMind
Operations research has always had an accessibility problem. Formulating a real-world scheduling, routing, or allocation problem as a mathematical program requires specialized expertise. Even when the formulation exists, asking what-if questions (“what happens if this warehouse closes?”) means modifying constraints and re-solving, a cycle that takes hours or days. OptiGuide breaks this loop by inserting a large-language model (LLM) between the decision-maker and the solver. The user asks a question in plain English, the LLM translates it into solver-ready code, the optimization engine executes, and the LLM summarizes the results in natural language. The design is deliberately privacy-preserving: proprietary data stays with the solver, never reaching the LLM. OptiMind, the companion model to OptiGuide, is a 20 billion-parameter reasoning small-language model, which is fine-tuned to formulate mixed-integer linear programs from natural-language business problems.
Data Formulator
Data visualization should be easy; it shouldn’t require you to become a programmer. You know what your data is saying; you just need a chart to tell that story in a visual way. Data Formulator is a concept-driven visualization tool that separates what you want to see from how to transform the data. It lets you describe data concepts in natural language or by example, and then bind them to visual channels. An AI agent handles the rest: reshaping tables, deriving new columns, writing transformation code, and rendering the chart from a library of 30 options. The result is a chart, produced in seconds, that would have taken a data scientist 20 minutes or more.