UMA Release Changelog

This page documents the release history of UMA models, including new features, improvements, and bug fixes.

Library compatibility — UMA 1.0 deprecation¶

UMA 1.0 checkpoints are no longer supported (their MoE include_self behavior diverges from later releases); loading one raises a RuntimeError. To use a 1.0 checkpoint, pip install 'fairchem-core<=2.21.0'.

Loading now requires a model_id to be defined. UMA 1.1 (which ships without one) is auto-tagged model_id = "UMA-1.1" in memory; UMA 1.2+ already carry model_id; a checkpoint with neither model_id nor backbone.model_version is rejected. When training, set model_id on the model config (e.g. model_id: UMA-1.2.1).

UMA-S-1.2.1-FT-D4rev and UMA-S-1.2.1-FT-r2SCAN — October 2026¶

• uma-s-1p2p1-ft-d4rev: omol task at the ωB97M-D4rev level of theory, fine-tuned on approximately 4 million OMol25 structures recomputed with ωB97M-D4rev labels.

• uma-s-1p2p1-ft-r2scan: omat task at the r²SCAN level of theory, trained on MAD-1.6.

See the UMA models page for usage, limitations and acknowledgements.

UMA Nitro and D3 dispersion — October 2026¶

• UMA Nitro is the default inference mode and is much faster than turbo for single-GPU MD and relaxations. It runs uma-s-1p2p1, uma-s-1p2p1-ft-d4rev and uma-s-1p2p1-ft-r2scan on Ampere, Hopper and Blackwell GPUs. Special thanks to Emanuele Arcelli and Max Zhdanov (for their FlashUMA work), Carole-Jean Wu, Christian Puhrsch, Zachary DeVito, Aditya Venkataraman, Michael Lazos, and Animesh Jain, all of whom contributed to and helped improve UMA Nitro. On a single NVIDIA H200, it reaches ~6 ns/day for 1,000 atoms in ASE MD and raw inference throughput equivalent to ~33 ns/day for 150 atoms, with support for systems up to 15,000 atoms.

• DFTD3Calculator adds GPU DFT-D3(BJ) energies, forces and analytic stresses to any ASE calculator, with pbe and r2scan presets. See Adding a DFT-D3(BJ) dispersion correction.

Library change — the default inference mode is nitro, and turbo is no longer available¶

• inference_settings="default", the default for get_predict_unit, load_predict_unit and FAIRChemCalculator.from_model_checkpoint, is now the nitro preset, which we recommend for MD and other standard workflows. It runs UMA Nitro for uma-s-1p2p1 (and the fine-tunes of it on Nitro’s checkpoint allowlist) on Ampere, Hopper and Blackwell GPUs, and multigpu otherwise, warning once that results are reduced precision.

• The previous default (full FP32 with merge_mole + compile) is now inference_settings="high-precision".

• Breaking: turbo is no longer available; please use nitro (the default). Its former settings are inference_settings="multigpu", which multi-worker inference (workers > 1) uses.

• The batch server’s ModelSpec and the torch-sim interface default to "batch".

• Fallback warnings name the settings that replace the requested ones.

• On Linux x86_64 and aarch64, nvidia-cutlass-dsl and cuda-python are now installed with fairchem-core.

UMA 1.2¶

Model Highlights¶

Available Tasks (7 total)¶

Large-Scale Inference & Molecular Dynamics¶

UMA 1.1¶

Available Tasks (5 total)¶

UMA 1.0¶

Available Tasks (5 total)¶

Key Features¶

• Mixture-of-Linear-Experts (MoLE) architecture for high parameter count with fast inference

• Equivariant GNN preserving energy conservation

• Multi-domain training on 500M+ DFT calculations

• ASE Calculator integration for seamless workflow adoption