Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableMeta Corporate News→ original

Meta releases UMA Nitro: an atomic model that found drug crystal forms without fine-tuning

Chemists first screen drug candidates on a computer, but precise atom-level calculations take a long time. Meta has released UMA Nitro, a faster version of its UMA neural network, which predicts how atoms behave. Working with Carnegie Mellon and UC Riverside, the model found many crystal forms of drugs without fine-tuning. Crystal form affects how quickly a tablet dissolves. This open model can serve as a first filter before the lab.

AI-processed from Meta Corporate News; edited by Hamidun News
Meta releases UMA Nitro: an atomic model that found drug crystal forms without fine-tuning
Source: Meta Corporate News. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceMeta Corporate News ↗
Editor-in-Chief
Key metric #1With the University of Chicago, UMA simulated entire enzymes atom by atom
Key metric #2High school chemistry teachers will get UMA Playground through Discovery Education
Key metric #3Data for the Digital Brain Project will grow from 15 000 to 25 000 hours

Before a new drug or material reaches a test tube, chemists run hundreds of candidates through computer simulations. That is cheaper than lab experiments, but precise atom-level calculations take a lot of time. This is the problem Meta is targeting. At the Open Source AI for Science summit in Menlo Park, the FAIR team (Meta's fundamental AI research division) released UMA Nitro, a faster version of its Universal Model for Atoms (UMA). It is a neural network that predicts how atoms interact in molecules and materials. And "without fine-tuning" means the model handled a new task without receiving additional examples for it.

Here is a simple example. A pharmacist wants to know which crystal forms a drug molecule can take. The form determines how quickly a tablet dissolves. Until now, answering that took heavy calculations or long series of lab experiments. Now a neural network can take over part of the work.

What was shown at the summit

The summit covered more than one model. Over two days, researchers from universities, hospitals, and industry discussed how open models, tools, and data help medicine, neuroscience, chemistry, materials science, and mathematics. Participants also worked in hands-on workshops alongside the authors of Meta's models.

For developers and founders building products close to science, the most concrete story is UMA-based chemistry. But the other announcements are also worth noting.

Drug crystals and enzymes

In a study with Carnegie Mellon and UC Riverside, UMA models found many crystal forms of drug molecules without additional training data. In practice, this matters: you do not need to build your own dataset and fine-tune the model for each molecule.

Together with the University of Chicago, Meta showed that UMA can simulate entire enzymes atom by atom. Enzymes are proteins that drive chemical reactions in the body. If researchers understand how they work, they can try to improve them for medicine, ecology, and energy. Meta is cautious here: it describes potential help for scientists, not finished discoveries.

The announcement gives no concrete figures for UMA Nitro's speed, so the speedup will have to be verified by independent tests.

From the lab to the classroom

UMA Playground is an interactive environment where users can build molecules, watch bonds break and form, and work through chemical reactions. Discovery Education announced that the tool is available to high school chemistry teachers, along with teaching materials. The platform reaches more than 4 million educators and 45 million students. RealChem, a college course from Arizona State and Carnegie Mellon used by almost 100 institutions, plans to add Playground to student assignments.

Brain, robots, and mathematics

Meta has expanded the Digital Brain Project, launched with the Rothschild Hospital Foundation and the University of Montreal. The target for open, de-identified brain data grew from 15 000 to 25 000 hours. The University of Pittsburgh showed an assistive robotics system built on DINOv3 and SAM 3. The team plans to demonstrate a robot that will help a person eat, and it is testing free-form voice commands.

In mathematics, the focus is on Lean, an open language in which a computer checks every logical step of a proof. This spares reviewers months or years of re-checking, whether the result comes from a colleague, a student, or an AI assistant. For this work, FAIR presented ATLAS, an open library that has formalized 26 textbooks and more than 45 000 computer-verified definitions and results.

What this means in practice

If you are building a startup in biotech, materials science, or chemical engineering, you can try UMA as a first-pass filter: weed out weak candidates before incurring lab costs. There are honest limitations, too. The crystal results come from a research project, not an industrial process, and the enzyme work remains a demonstration. The announcement gives no accuracy figures, so before adopting the model, test it on your own data and compare it with classical calculations. Still, the open models and code provide a base layer on which you can build tools instead of writing simulators from scratch.

Story

Key Questions Answered

What was shown at the summit?

The summit covered more than one model. Over two days, researchers from universities, hospitals, and industry discussed how open models, tools, and data help medicine, neuroscience, chemistry, materials science, and mathematics. Participants also worked in hands-on workshops alongside the authors of Meta's models. For developers and founders building products close to science, the most concrete story is UMA-based chemistry. But the other announcements are also worth noting.

What are the key details on Drug crystals and enzymes?

In a study with Carnegie Mellon and UC Riverside, UMA models found many crystal forms of drug molecules without additional training data. In practice, this matters: you do not need to build your own dataset and fine-tune the model for each molecule. Together with the University of Chicago, Meta showed that UMA can simulate entire enzymes atom by atom. Enzymes are proteins that drive chemical reactions in the body.

What are the key details on From the lab to the classroom?

UMA Playground is an interactive environment where users can build molecules, watch bonds break and form, and work through chemical reactions. Discovery Education announced that the tool is available to high school chemistry teachers, along with teaching materials. The platform reaches more than 4 million educators and 45 million students. RealChem, a college course from Arizona State and Carnegie Mellon used by almost 100 institutions, plans to add Playground to student assignments.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…