Microsoft Research announced Skala 1.1 on August 20, 2026, along with a set of integrations meant to move its machine-learned exchange-correlation functional out of a research demo and into the software chemists actually run. Density functional theory is the workhorse of computational chemistry and materials science, and its accuracy is bounded by the quality of the approximate functional at its core. Skala replaces that hand-derived approximation with a learned one.
The headline number is a weighted average error of 2.8 kcal/mol on GMTKN55, the standard general-purpose main-group thermochemistry benchmark, with Skala 1.1 ranking first in 32 of the benchmark’s 55 categories. Microsoft trained the new version on 2.5 times more data than its predecessor and says it outperforms today’s leading range-separated hybrid functionals while keeping the runtime profile of a semi-local functional. On the cost side, the company reports GPU performance comparable to r2SCAN, with CPU overhead diminishing above roughly 300 orbitals and effectively disappearing for molecules of 20 to 30 atoms or more. That combination - hybrid-class accuracy at meta-GGA cost - is the whole argument for the approach.
Access is the other half of the announcement. Skala is now fully integrated into CP2K, done in collaboration with Thomas Kuehne’s Center for Advanced Systems Understanding, with integrations in progress for Psi4, FHI-aims, ORCA and VASP. There is an open-source release through GPU4PySCF with ASE integration, and Microsoft introduced what it calls a living benchmark plus a performance harness to track successive Skala releases across hardware platforms and implementations - an unusual and welcome commitment, since learned functionals are exactly the kind of artifact whose reported accuracy is hard to compare across versions.
For a technical leader in pharma, chemicals, batteries or semiconductors, the question is not whether the benchmark number is impressive but whether it transfers. GMTKN55 is main-group thermochemistry; transition metals, extended solids and reaction barriers are where DFT most often disappoints in practice, and Microsoft’s post does not claim those are solved. The practical significance is that a learned functional is now reachable from the mainstream DFT codes rather than requiring a bespoke pipeline, which lowers the cost of running the only test that matters: whether it holds up on your own chemistry.