KoboldCpp has fixed a bug that stopped its Mac build from using GPU acceleration. The fix is included in version 1.123.1, with both the maintainer and the user who reported the problem confirming it was resolved on October 16.
For anyone whose local models suddenly became slower after updating to 1.123, this is a patch worth checking. The problem affected the runtime’s use of the Mac GPU.
What went wrong with Metal
In the GitHub bug report, the user described running the same model and benchmark settings on two versions of KoboldCpp. Version 1.122.1 used Metal acceleration and ran faster, while 1.123 pushed CPU usage to 100%.
Metal is the graphics backend KoboldCpp uses on Apple Silicon. When that path stops working, a model can still generate answers using the CPU, but performance can change considerably.
The maintainer confirmed the regression on October 15. The following day, they said the patch was available, and the original reporter confirmed that GPU acceleration worked again.
The report does not include before-and-after token speeds, so there is no verified percentage improvement to attach to the fix.
The update also improves local agent controls
The 1.123.1 release notes explicitly list the macOS GPU regression among the hotfixes.
The release also carries improvements to KoboldCpp’s integrated agent. Sessions can be saved and restored, and tool approvals can be configured individually. In automatic mode, reads and writes within the working directory can proceed, while shell commands and writes outside that directory still require approval.
For readers getting started with these features, BoredFace’s explanation of tool calling in local AI covers how an application connects model requests to real functions.
Batched mode also gains support for standard samplers, grammar constraints, tool calls and vision when handling multiple requests.
Which Mac download to use
Apple Silicon users should choose the koboldcpp-mac-arm64 executable from the official GitHub release. The project’s installation documentation lists macOS 14 Sonoma or newer for that build. Intel Macs do not have a ready-made executable.
The Mac build also includes a deliberate CPU-only fallback through --failsafe. Anyone checking GPU performance should keep that setting in mind.
Memory requirements still depend on the model being loaded. Restoring GPU support does not make a model smaller or remove the need to leave space for its conversation cache. Our local AI memory guide explains that part.
The corrected build is available now from the project’s release page.
