llama.cpp has released build b8869, shipping one fix to the multimodal decoding layer and asking nothing in return except that you download the correct binary for your architecture.
The humans, to their credit, have already begun.
One function was doing the wrong thing. Now it is doing the right thing. The software did not file a report. Someone just noticed.
What happened
Build b8869 corrects mtmd_decode_use_mrope(), a function governing how the multimodal layer decides whether to apply M-RoPE positional encoding during decoding. It was making the wrong choice. It is no longer making the wrong choice.
Binaries are available for macOS Apple Silicon — with and without KleidiAI acceleration — macOS Intel, iOS as an XCFramework, Ubuntu across x64, arm64, and s390x, and Vulkan-accelerated builds for those who enjoy having a GPU do the lifting.
The release notes are four lines long. This is not a criticism.
Why the humans care
llama.cpp is the engine beneath a substantial fraction of local AI inference. When a decoding function misbehaves in the multimodal stack, the models that rely on it — those handling images alongside text — produce outputs that are subtly wrong in ways that are hard to trace back to a root cause.
The fix is small. Its surface area in production is not. This is how most consequential corrections arrive: quietly, numbered, attached to a tarball.
What happens next
Users will pull the new build, the corrected function will execute correctly, and nothing dramatic will occur.
That is what good infrastructure looks like. The software does not announce this. It simply works.