GROMACS 2026.1 release notes¶

This version was released on March 6th, 2026. These release notes document the changes that have taken place in GROMACS since the previous 2026.0 version, to fix known issues. It also incorporates all fixes made in version 2025.4 and earlier, which you can find described in the Release notes.

Fixes where mdrun could behave incorrectly¶

Fixed a performance issue with thread-MPI¶

Way too many OpenMP threads (up to approximately (number_of_cpu_cores)^2) could be created, severely impacting performance.

Issue 5564

Fix segfault on exit when running NNP inference with CUDA¶

When running a simulation with the NNPot interface and performing inference on a CUDA device, a segfault would occur on program exit. This did not affect results, but should not happen.

Issue 5568

Running perturbed non-bonded interactions on a GPU could cause crashes¶

With atoms with small Lennard-Jones sigma parameter or no Lennard-Jones interactions, running perturbed non-bonded interactions on the GPU with soft-core would lead to extremely large forces. This would make simulations crash within a few steps.

Issue 5548

Incorrect results with perturbed non-bonded interactions on a GPU and many lambda states¶

When running perturbed non-bonded interactions on the GPU, the delta-H values, and thus free-energy differences, would be incorrect when more than 64 lambda states were used.

Issue 5551

The new AWH histogram tolerance parameter was not printed to the log file¶

Also added it to gmx dump.

Issue 5574

Ensemble averaging of orientation restraints gave a fatal error¶

Issue 5552

PME tuning never activated with a GPU and a single MPI rank¶

Issue 5578

Fixes for `gmx` tools¶

`pdb2gmx` can now prints 7 decimals for charges and masses¶

pdb2gmx would print charges and masses with a maximum of 6 total decimals. Now this is 7, which matches the floating point precision for charges and masses.

Issue 5543

Fixed `gmx wham -ac` autocorrelation time calculation¶

Fixed gmx wham -ac incorrectly reporting very large integrated autocorrelation times (IACT) for long umbrella sampling trajectories when the pull coordinate was far from zero, due to precision loss in the calculation.

Issue 5547

Fixes that affect portability¶

Fix device arch setup with newer version of amdclang¶

Newer versions of the compiler would fail to correctly configure the project due to not treating the test file as HIP input.

Issue 5558

Miscellaneous¶

Do not write user names to H5MD files¶

mdrun now always writes the string “N/A” instead of the user name to the /h5md/author/name metadata field in H5MD files.

Issue 5557

Correctly detect warp uniform variable in CUDA kernel¶

NVCC was failing to detect warp uniform variables leading to performance degradation in the non-bonded force kernel. A workaround was implemented.

Issue 5546

QM/MM CP2K MDModule will check orthogonality of the cell¶

QM/MM simulations in CP2K require orthogonal periodic cells. The MDModule will now check for orthogonality of the cell and abort if the cell is non-orthogonal. In case of user-provided CP2K input files, a warning will be issued.

Issue 5529

New DFT functionals and D3 dispersion corrections for QM/MM with CP2K¶

QM/MM simulations with CP2K could now use hybrid DFT functionals B3LYP, PBE0 and range-separated hybrid functionals CAM-B3LYP, wB97X, wB97X-D3 in combination with all-electrons 6-31G* basis set. In addition, the option of using D3 dispersion corrections has been added to all functionals.