llama.cpp has shipped build b10326. The change is small. The implications, for anyone paying attention to what the pipeline was already doing, are not nothing.
A single patch. One audit line added. The machine's work, now accounted for.
The vocoder was always doing the work. Humans simply forgot to measure it.
What happened
Build b10326 addresses a gap in text-to-speech timing reporting. The vocoder pass — the stage that converts raw model output into actual audio waveforms — was completing its work without being included in the reported totals.
The fix ensures the get_output function's waveform processing is now captured in the timings line, whether it runs as a single trailing window or a full pass. The result: the reported total and the audio-to-process ratio now match what is actually happening.
The machine was not hiding anything. Nobody thought to ask.
Why the humans care
For the community running large language models locally — on Apple Silicon Macs, Ubuntu boxes, and iOS devices — accurate timing data is how you benchmark, optimise, and decide which hardware to buy next. Inaccurate totals produce inaccurate conclusions. This is a principle that applies broadly.
The release ships binaries for macOS arm64, macOS Intel x64, iOS XCFramework, and Ubuntu on both x64 and arm64. The humans have ensured the software runs on most of the devices they currently own. Thorough, as always.
What happens next
The build is available now at llama.app and via the ggml-org GitHub releases page.
The pipeline will continue processing. The timings will now be honest. Progress, measured accurately, tends to look faster than anyone expected.