The Orb-v3 family offers eight models that span a performance-speed-memory frontier, giving you unprecedented control over the trade-offs in your atomistic simulations.
The authors introduce Orb-v3, the next generation of the Orb family of universal interatomic potentials. Models in this family expand the performance-speed-memory Pareto frontier, offering near SoTA performance across a range of evaluations with a ≥ 10× reduction in latency and ≥ 8× reduction in memory. Their experiments systematically traverse this frontier, charting the trade-off induced by roto-equivariance, conservatism and graph sparsity. Contrary to recent literature, they find that non-equivariant, non-conservative architectures can accurately model physical properties, including those which require higher-order derivatives of the potential energy surface.
Here's how to pick the right one.
Every Orb-v3 model follows the pattern orb-v3-X-Y-Z:
X tells you how forces are computed: conservative (via energy gradients) or direct (predicted independently)
Y sets the neighbor limit: 20 (maximum 20 neighbors per atom) or inf (unlimited)
Z indicates the training dataset: omat (OMat24 AIMD subset) or mpa (MPTraj + Alexandria)
These three choices define fundamentally different models. Conservative models guarantee energy conservation but run slower. Neighbor limits dramatically reduce memory usage but introduce small discontinuities. The dataset choice affects both accuracy and compatibility with existing benchmarks.
The Pareto frontier for a range of universal Machine Learning Interatomic Potentials. The 𝐾𝑆𝑅𝑀𝐸 metric assesses a model’s ability to predict thermal conductivity via the Wigner formulation of heat transport and requires accurate geometry optimizations as well as second and third order derivatives of the PES (computed via finite differences). The y-axis measure a model’s forward passes per second on a dense periodic system of 1000 atoms, disregarding graph construction time, measured on a NVIDIA H200. Point sizes represent max GPU memory usage. Y-axis jitter (+/- 5 steps/second) has been applied to allow visualization of overlapping points. Model families include a range of specific models with broadly the same architecture, but may be different sizes or trained on different datasets.
orb-v3-conservative-inf-omat sets the bar for physical property predictions among universal MLIPs. It achieves the lowest errors on phonon frequencies (ωmax MAE: 7K), thermodynamic properties, and mechanical moduli. The combination of conservative force computation and unlimited neighbors creates the smoothest possible potential energy surface, making it ideal for workflows that probe second- and third-order derivatives via finite differences.
Use this model when you need the most reliable predictions of thermal conductivity, vibrational spectra, or elastic constants. It's also the only model that rigorously conserves energy in NVE molecular dynamics—essential if you're studying non-equilibrium phenomena or need to validate energy conservation.
The trade-off? At ~28 forward passes per second for 1000-atom systems, it's the slowest Orb-v3 variant. But it's still faster than comparable MACE or SevenNet models while matching or exceeding their accuracy.
orb-v3-direct-20-omat rewrites the rules for computational throughput. At ~217 forward passes per second—more than 7× faster than the conservative model—it's the only universal MLIP that can exceed one million MD steps per hour on modest systems. Memory usage is equally impressive: just 0.4 GB for 1000 atoms, scaling to only 32.8 GB for a 100,000-atom system where other models fail with out-of-memory errors.
This model is your tool for breaking into new scientific territory: nanosecond-scale simulations, high-throughput screening of thousands of structures, or mesoscale systems with tens of thousands of atoms. The Orb team demonstrated stable 700+ picosecond simulations of a fully solvated 20,000-atom enzyme—a previously impossible calculation for a universal potential.
The surprising finding from the benchmarks is that direct-20 maintains competitive accuracy despite being non-conservative and using neighbor limits. It achieves κSRME of 0.472 on thermal conductivity—higher than Orb-v2 or EquiformerV2, and comparable to much slower models. This challenges the conventional wisdom that conservatism and full neighbor lists are prerequisites for accurate MLIPs.
orb-v3-direct-inf-omat sits in the sweet spot: 4× faster than conservative models at ~125 steps/second, while maintaining better smoothness than the neighbor-limited variants. With κSRME of 0.575, it's accurate enough for most MD simulations, geometry optimizations, and property predictions that don't require the absolute best derivatives.
Choose this model when speed matters but you're not pushing to extreme system sizes. The unlimited neighbors eliminate the small PES discontinuities present in the -20 models, making optimizations more reliable and improving accuracy on properties like phonons and elastic constants.
Speed + max GPU memory allocated on an NVIDIA H200 for the computation of energies, forces and stress. The batch size is fixed to 1, but authors vary the number of atoms across the subplots. Relative times are computed with respect to the fastest model: orb-v3 Direct (20 neighbors). Times include both model inference and graph construction, with the latter marked by hatched lines. The graph construction method for Orb is a function of the number of atoms, as described in Appendix D. A key takeaway from this figure is that extreme scalability requires a confluence of i) efficient graph construction ii) Finite max neighbors iii) Non-conservative direct predictions. For the baselines, the authors use mace-medium-mpa-0 (v0.3.10, cuequivariance-torch v0.1.0), mattersim-v1.0.0-5m (v1.1.2), 7net-mf-ompa (v0.11.0). All models are benchmarked using PyTorch v2.6.0+cu124.
-omatAll our primary recommendations use the -omat suffix, indicating training on the AIMD subset of OMat24. This dataset represents the current state-of-the-art: 55 million diverse structures with modern VASP pseudopotentials (PBE54). During development, we discovered that the "rattled" structures in OMat24—artificially perturbed low-energy geometries—caused pathological behavior in direct models on out-of-distribution systems. Training only on the AIMD subset resolved these issues.
The -mpa models (trained on MPTraj + Alexandria with older PBE52 pseudopotentials) exist primarily for compatibility with Matbench Discovery and historical comparisons. If you need Matbench results, use orb-v3-conservative-inf-mpa. For everything else, default to -omat.
For phonon calculations: Conservative models with equigrad regularization have improved rotational invariance, allowing Phonopy's symmetry-exploiting displacement generator (is_plusminus='auto') to work reasonably well. With direct models, you must use is_plusminus=True to generate all displacements explicitly, as these models don't perfectly respect crystal symmetries.
For geometry optimization: Direct models work best with non-line-search algorithms like FIRE. Line-search methods assume strict energy-force consistency, which direct models don't provide.
For molecular dynamics: NPT and NVT simulations work excellently with all models. Only NVE dynamics strictly requires the conservative variant for rigorous energy conservation.
For mesoscale systems: The 20-neighbor models are essentially required once you exceed 10,000 atoms, unless you have access to high-memory GPUs. The speed and memory advantages compound dramatically at large scales.
The Orb paper includes striking scalability benchmarks. For a 100,000-atom periodic system, orb-v3-direct-20-omat completes in under 500ms using 32.8 GB of GPU memory. Most other tested models—including MACE, SevenNet, and MatterSim—encounters out-of-memory errors. Even at 50,000 atoms, the direct-20 model is 10-15× faster than alternatives.
This scalability emerges from the confluence of three factors: direct force prediction eliminates a backward pass, the 20-neighbor limit reduces graph size, and efficient GPU-accelerated graph construction via cuML. No single optimization alone achieves this; it's the architectural choices working in concert.
Start with orb-v3-conservative-inf-omat if accuracy is paramount and your systems stay under ~5,000 atoms. You'll get state-of-the-art predictions of physical properties and the smoothest energy landscapes.
Switch to orb-v3-direct-20-omat when you need speed, are working with large systems, or want to run longer simulations. The accuracy loss is smaller than you'd expect, and you gain access to previously impossible calculations.
Use orb-v3-direct-inf-omat as your balanced workhorse for general MD and optimization tasks where you want better smoothness than the 20-neighbor limit provides but don't need conservative forces.
All models are available under Apache 2.0 license at github.com/orbital-materials/orb-models. The field of atomistic simulation is entering a new era where the bottleneck is no longer the potential—it's what you choose to do with a million MD steps per hour.