UMA: A Family of Universal Models for Atoms: Meta Open-Sources Universal Atomic Simulation Model to Accelerate Materials and Chemistry Calculations
In June 2025, Meta FAIR released the UMA model family, trained on 500 million 3D atomic structures, using a mixture of linear experts architecture. The small/medium models have 1.4B parameters but only about 50M active parameters. Without fine-tuning, they achieve or exceed the performance of specialized models in multiple domains including molecules, materials, and catalysts. Code, weights, and data have been open-sourced.
UMA is the first universal atomic simulation foundation model covering multiple chemical domains such as molecules, materials, and catalysts. Through large-scale data (500 million structures) and a mixture of linear experts architecture, it significantly improves accuracy while maintaining inference speed. Its zero-shot generalization capability challenges the traditional 'specialized model' paradigm, potentially unifying AI toolchains for computational chemistry and materials science, lowering R&D barriers.
UMA adopts a mixture of linear experts (MoLE) architecture, activating only about 50M parameters per atomic structure (total 1.4B parameters), achieving a balance between high capacity and efficiency. Training data covers 500 million 3D structures from molecules, materials, catalysts, etc., and model capacity and data scale are optimized based on empirical scaling laws. Evaluations show that UMA-medium achieves or exceeds the accuracy of specialized models on benchmarks such as QM9, MD17, and OC20 without fine-tuning. Its architecture avoids the computational bottlenecks of traditional graph neural networks, supporting parallel processing of batch atomic structures.
UMA will accelerate atomic-scale simulations in drug discovery, energy storage, semiconductor manufacturing, and other fields. Traditional DFT calculations take hours, while UMA can provide approximate accuracy in milliseconds, enabling high-throughput virtual screening. Meta's open-sourcing of model weights and code may give rise to materials design SaaS platforms, lowering the barrier for SMEs to use AI simulations.
It is recommended that materials/chemistry R&D teams immediately evaluate the potential of UMA-medium to replace existing computational pipelines, prioritizing its use for high-throughput virtual screening and candidate pre-filtering. Invest in commercial tools derived from Meta's open-source ecosystem (e.g., UMA API services) and specialized fine-tuned model services based on UMA.
Focus on UMA's deployment cost (GPU requirements) and accuracy boundaries (e.g., transition states, rare earth element systems) in industrial molecular dynamics simulations. If the community can reproduce its zero-shot generalization capability, it will promote AI-first R&D workflows in more fields (e.g., catalysis, batteries). Be cautious of data bias affecting rare systems.