The news is not just an accuracy upgrade. Skala 1.1, Microsoft Research's deep-learning exchange-correlation functional, is moving into the codes that the computational chemistry community uses every day on its own clusters and workstations. The team has made Skala available in CP2K and is working on integration into Psi4, FHI-aims, ORCA, and VASP. In parallel, it is publishing a living benchmark that will track performance over time.

The technical leap is concrete: trained on 2.5 times more data than the first public version, the model lowers the weighted average error to 2.8 kcal/mol on GMTKN55. According to Microsoft Research, it surpasses the best range-separated global hybrid functionals while maintaining the computational cost of a meta-GGA. In practice, simulations that previously required more expensive hybrid methods can now run with budgets similar to those of semi-local functionals.

The performance picture is not secondary. In the published comparison, on GPU Skala 1.1 has the same cost as r2SCAN, while the B3LYP and M06-2X hybrids become more expensive beyond roughly 1000 orbitals. On CPU the overhead relative to the other functionals disappears for systems above roughly 300 orbitals. This curve matters for infrastructure planning: the marginal cost of the accuracy jump falls just as systems become large, exactly when traditional calculation tends to weigh more.

A living benchmark for local validation

Validation across implementations is another step that separates a scientific announcement from an adoptable tool. With the CASUS team, Microsoft Research built a test suite for CP2K and verified agreement with the PySCF implementation within 0.1 kcal/mol MAD on a representative subset of GMTKN55. This is not a formal detail: for those running simulations on local infrastructure, reproducibility across different codes is often the real bottleneck before replacing an established functional.

The living benchmark adds a level of transparency that is still rare in applied AI. Instead of a single snapshot, Microsoft Research is publishing a harness and a report updated as optimizations arrive. This creates an incentive for maintainers of individual packages to compare on common ground and makes visible the effect of future releases on different hardware platforms.

The signal for on-premise deployment

Skala is not a Large Language Model, but its distribution path has direct implications for those managing scientific workloads on local infrastructure. It is not being distributed as a closed cloud service or as a separate API: it is being integrated into the open-source and commercial codes that run locally. On one hand, this avoids lock-in to a single package; on the other, it shifts adoption, testing, and maintenance work onto the communities of individual software packages. Research groups using CP2K, ORCA, or VASP can adopt the model without moving data outside their own systems, a relevant constraint in sectors where molecular data are sensitive or covered by confidentiality agreements.

Microsoft Research's choice to invest in multiple integrations, rather than a single proprietary platform, signals a broader direction: the value of a scientific model consolidates when it enters existing workflows, not when it remains confined to a separate environment. For those evaluating on-premise deployment, there are trade-offs among continuous updates, validation costs, and infrastructure control; AI-RADAR offers analytical frameworks on /llm-onpremise to place them in a wider picture.